RUTCOR

RESEARCH
R E P O R T

COUNTRY RISK RATINGS:

STATISTICAL AND
COMBINATORIAL NON-
RECURSIVE MODELS

P.L. H A M M E R a A. Kogan b M.A. Lejeune c

RRR 8-2004, MARCH 2004

RUTCOR
Rutgers Center for
Operations Research
Rutgers University
640 Bartholomew Road
Piscataway, New Jersey
08854-8003
Telephone: 732-445-3804 a
hammer@rutcor.rutgers.edu , Tel.: +1 7324454812 ; Fax: +1 7324455472
b
Telefax: 732-445-5472 kogan@rutcor.rutgers.edu
c
Email: rrr@rutcor.rutgers.edu mlejeune@andromeda.rutgers.edu
a,b,c
http://rutcor.rutgers.edu/~rrr
RUTCOR - Rutgers University Center for Operations Research, Piscataway, NJ, USA
a,b
Rutgers Business School, Rutgers University, Newark-New Brunswick, NJ, USA
 RUTCOR RESEARCH REPORT
RRR 8-2004, MARCH 2004

COUNTRY RISK RATINGS:

STATISTICAL AND COMBINATORIAL
NON-RECURSIVE MODELS

P.L. Hammer A. Kogan M.A. Lejeune

Abstract. The central objective of this paper is to develop transparent, consistent, self-
contained, and stable country risk rating systems, closely approximating the country risk ratings
provided by a major rating agency (Standard & Poor). We propose two models that achieve the
stated objectives, the first one utilizing the classical econometric technique of multiple linear
regression, and the second one using the combinatorial-logical technique of Logical Analysis of
Data. The proposed models use economic-financial and political variables, and are non-
recursive (i.e., they do not rely on the previous years’ ratings). The accuracy of the proposed
models’ predictions, measured by their correlation coefficients with Standard and Poor’s
ratings, and confirmed by k-folding cross-validation, exceeds 95%. The stability of the
constructed non-recursive models is shown in three ways: by the correlation of the predictions
with those of other agencies (Moody’s and The Institutional Investor), by predicting 1999
ratings using the non-recursive models derived from the 1998 dataset applied to the 1999 data,
and by successfully predicting the ratings of several previously non-rated countries. The
confidence in the results and in the validity of both models is strongly reinforced by the fact that
the traditional linear regression model and the qualitatively different combinatorial-logical
model produce almost identical results.

Acknowledgements: The authors express the appreciation to Dr. Sorin Alexe for his invaluable help in
the execution of computational experiments with LAD.
1 Country Risk, Country Risk Ratings and Objectives of the
Paper

1.1 Country risk, country risk ratings and their importance

The globalization of the world economies, and in particular the internationalization of
financial markets in the last decades, have dramatically expanded and diversified investment
possibilities, leading to numerous new opportunities, accompanied by new risks. Consequently,
there has been growing interest in obtaining reliable estimates of the risk of investing in different
countries. These concerns have led to the development of the concept of country risk, and even to
the regular publication of country risk ratings by various agencies. The importance of ratings has
been magnified by the recommendations addressed in the Basel Capital Accord (2001), that
pinpoints the role of agencies’ ratings for the assessment of credit risk.
Different definitions have been proposed for country risk, i.e. for the risk that a country
defaults on its obligations. The existing literature on the topic recognizes both financial/economic
and political components of country risk. According to the degree to which some of these
components are emphasized, country risk is viewed either from the financial/economic
perspective only, or from the combined financial/economic and political perspectives.
There are two basic approaches to the interpretation of the reasons for defaulting. The debt-
service capacity approach focuses on the deterioration of solvency of a country, which prevents it
from fulfilling its commitments. For instance, Bourke and Shanmugam (1990) define country risk
as “the risk that a country will be unable to service its external debt due to an inability to generate
sufficient foreign exchange”. Within this framework, country risk is viewed as a function of
various financial and economic country parameters. The cost-benefit approach views a default on
commitments or a rescheduling of debt as a deliberate choice of the country, which may prefer
this alternative over repayment, in spite of its possible long-term negative effects (e.g. the
country’s exclusion from certain capital markets (Reinhart,2002), reputation damage). Since the
deliberate decision to default results from a political process, political country parameters are
included in this type of country risk modeling, along with the financial and economic ones. This
approach is strongly recommended by Brewer and Rivoli (1990, 1997) as well as Citron and
Neckelburg (1987), who emphasize the impact of the political stability indicator on country risk
ratings.
In response to the increased demand for the evaluation of creditworthiness, several agencies
such as Moody’s, Standard & Poor, Fitch, The Institutional Investor, Euromoney, Dun &
Bradstreet, etc. have developed expertise in estimating country risk. These estimates are
presented in the form of ratings, or scores, and are generally viewed as indicative of possible
future default. Haque et al. (1996) define country credit risk ratings compiled by commercial
sources as an attempt “to estimate country-specific risks, particularly the probability that a
country will default on its debt-servicing obligations”. Sovereign ratings can be viewed as the
probability that a borrowing country will fail to pay back.
Country (or sovereign) risk ratings impact countries in a number of ways. The primary
significance of ratings is due to their influence on the interest rates at which countries can obtain
credit on the international financial markets: the higher the ratings (i.e., the lower the risk of
RRR 08-2004 PAGE 1

default) the lower the interest rate. Following its sovereign rating downgrade , Japan’s borrowing
became more expensive as interest rates have increased, reflecting the higher chance of default ,
which deteriorates even more the situation of the heavily indebted Japanese government and
economy.
Second, sovereign ratings also influence credit ratings of national banks and companies, and
affect their attractiveness to foreign investors. Ferri et al. (2001) call sovereign ratings the “pivot
of all other country’s ratings”. Similarly, Erb et al. (1995a) underline that raters have historically
shown a reluctance to give a company a higher credit rating than that of the sovereign where the
company operates. For example, after Moody’s downgraded Japan in November 1998 (from Aaa
to Aa1), all other Aaa Japan issuers have been downgraded (Jüttner and McCarthy, 2000). This
led sovereign ratings to be named “sovereign credit risk ceilings”.
Third, institutional investors are sometimes contractually restricted on the degree of risk they
can assume, implying in particular that they cannot invest in debt rated below a prescribed level.
Ferri et al. (2001) refine this analysis, pointing out the contrast between the ratings of banks
operating in high- and low-income countries, and show that ratings of banks operating in low-
income countries are significantly affected by variations in sovereign ratings, while the ratings of
banks operating in high-income countries do not seem to depend significantly on country ratings.
Similarly, Kaminsky and Schmukler (2002) as well as Larrain et al. (1997) note that sovereign
ratings are crucial for developing economies, which have a very high sensitivity to rating
announcements.

1.2 Critiques of present rating systems

The purpose of ratings is that of compressing a variety of information about a country into a
single parameter which can be easily understood, and therefore conveniently used in a decision
making process involving comparisons between different countries. Consequently, ratings
provide aggregations of diverse indicators into a single metric and can be viewed as a kind of
“commensuration” (Kunczik, 2000). The interpretation of ratings is complicated by the
heterogeneity of indicators (political stability, inflation, etc.) which may have been used in
deriving them.

Comprehensibility: The country risk ratings published by different agencies appear as outputs of
“black boxes”, the real content and meaning of which are unexplained and hard to understand,
since rating agencies specify neither the factors which are taken into consideration in determining
their ratings, nor the “rules of compression” of multiple factors into a single rating. This raised
the discontent of Japan’s Prime Minister, Junichiro Koizumi, who was “railed at being rated in
the same neighborhood as African countries to which Japan is providing assistance” . Officials of
Japan’s Ministry of Finance added that big rating agencies are “making unfair qualitative
judgments” , while Moody’s denied and claimed that the motives for the downgrade lie in the
“increased debt load” of Japan. In view of such controversy, uncovering both the factors which
are taken into account by these black boxes, and the mechanisms of deriving ratings, are essential
for ascertaining the consistency of a country rating system.

Unknown factors: It is generally assumed that ratings are obtained by aggregating

economic/financial and/or political variables. Clearly, the main objective of any country risk
rating system is to represent the creditworthiness of countries, i.e., their capacity to pay off loans.

1
PAGE 2 RRR 08-2004

It is not clear however which ones of the many possible factors do actually influence the payback
capacity of a country. This question is subject to different analyses. Haque et al. (1998) claim that
it is sufficient to restrict the scope of analysis to economic/financial factors only, while others
(Brewer and Rivoli,1990) claim that both economic/financial and political factors impact country
risk ratings.

Rating failures: Some recent failures have challenged the trustworthiness of country risk ratings
(Reinhart, 2002, Levich et al., 2002) . Criticisms directed towards ratings institutions have been
especially intense after the Tequila and the Asian crises. Indeed, the tequila crisis in Mexico
(1994-95) had not been preceded by a rating downgrade, implying that either the crisis was not
predicted, or that its significance was overlooked. Similar observations apply to the Asian crisis
(1997-99): Ftich admitted that “it and its larger rivals Standard’s & Poor and Moody’s Investors
Services of the US had largely failed to predict the recent turmoil in Asia” . On the other hand,
rating agencies have been more insightful in anticipating other crises, e.g. in Russia (1998),

• Regional bias: Diverse explanations have been provided for the failure of rating
Brazil (1998) and Argentina (2001).

agencies to signal crisis emergencies in various countries. There are claims that certain
rating agencies favor certain regions. For instance, Haque et al. (1997) note that
Euromoney usually gives higher ratings to Asian and European countries than to Latin
or Caribbean countries, while the Institutional Investor is more generous to Asian and

• Latency: Another criticism lies in the time taken by the rating agencies to react to new
European countries than to African ones.

facts (e.g., according to The Economist , “rating agencies may have been too slow to

• Overreactions: The IMF criticizes rating agencies claiming that they reacted in panic
downgrade Japan. Markets have already moved ahead of them”).

during the Asian crisis. After they had missed to predict the Asian crisis, they reacted by
harshly downgrading countries such as Thailand or South Korea, thus accelerating the
flight of capital. In this and other situations, rating agencies gave the impression of
overreacting (Figure 1) instead of being a stabilizing force.
Figure 1: Precrisis and postcrisis rating of countries
The Institutional Investor's ratings

80 Precrisis 80 Precrisis
70 Postcrisis 70 Postcrisis
Euromoney's ratings

60 60

50 50

40 40

30 30

20 20

10 10

0 0
Thailand Korea Indonesia Thailand Korea Indonesia

It appears that the objectivity and reliability of country risk ratings is questionable, mainly
because of human intervention and conflicting goals and/or interests.
RRR 08-2004 PAGE 3

• Negative impact of rating changes: It is reported that the hesitation or reluctance of

raters to downgrade a country stems from the fact that a downgrade announcement can
precipitate a country into crisis. During the Asian crisis, the rating agencies arouse the
discontent of the Malaysian Prime Minister, Dr Mahathir bin Mohamad, who
condemned them and charged them with rendering the crisis even more acute. “The
rating agencies, when we have a need to borrow money, they immediately downgraded
us so that it will cost us 15% to borrow money. They stop us completely from
borrowing money” (1999)1. Along the same line, Reisen and Von Maltzan (1999) claim
that such a sharp downgrade impeded “commercial banks to issue letters of credit,
forced investors to offload Asian assets to maintain portfolios in investment-grade
securities”. They argue that rating agencies lagging behind rather than anticipating the
state of financial markets reinforce positive expectations and capital inflows when they

•
upgrade countries and intensify outflows of capital and crisis when they downgrade.
Conflicts of interest: An even more pointed criticism is that raters, having started
charging fees to rated countries, can be suspected of reluctance to downgrade them,
because of the possibility of jeopardizing their income sources. This is claimed, for
example, by Tom McGuire, an executive vice-president of Moody’s, who states that
“the pressure from fee-paying issuers for higher ratings must always be in a delicate
balance with the agencies’ need to retain credibility among investors”2. The necessity to
please the payers of the ratings, investors as well as issuers, lead to what Robert
Grossman, the chief credit officer at the rating agency Fitch, calls “a tendency we do
with investors – rating committees, outlooks, meetings, then the press release, all to
soften the blow of the rating change”3. Studying the rating transitions, Altman and
Saunders (1998) notice that a downgrade in the rating of a country is regularly followed
by further downward adjustments. The explanation given by Altman and Saunders is
that agencies gradually downgrade the rating of a country, since they do not want to hurt
the country, which is also their client. Kunczik (2001) note that the IMF (1999) fears the
danger that “issuers and intermediaries could be encouraged to engage in rating
shopping – a process in which the issuer searches for the least expensive and/or least
demanding rating”.

The problems described above will become more acute as the role of ratings increases. Indeed,
the Basel Accord will intensify the pressure on countries to obtain high ratings, potentially
leading to a switch from rating shopping to rating fraud. For instance, Pakistan has been forced to
pay back $55 million credits to the IMF because of budget falsification, the blame being put on
the former Prime Minister Nawaz Sharif, accused of having falsified the budget deficit. Similarly,
Ukraine has been proven to have reported misleading data on its reserves in foreign exchanges,
attempting to obtain IMF credits. Kunczik (2001) says that “it is only a question of time when
firms will specialize in rating advising for sovereigns”.

1
This article appeared in the February 19, 1999 issue Executive Intelligence Review.
Interview: Datuk Seri Dr. Mahathir bin Mohamad Malaysian Prime Minister: `We had to decide things for ourselves'.
On January 22, 1999, Gail G. Billington of EIR's Asia Desk and Dino de Paoli of the Schiller Institute were given the
opportunity to interview Datuk Seri Dr. Mahathir bin Mohamad, Prime Minister of Malaysia.
2
The Economist, July 15, 1995, 62
3
Euromoney, January 2002, 38, “Investors turn cool on the rating game”

3
PAGE 4 RRR 08-2004

1.3 Can yield spreads replace country risk ratings?

To mitigate the problems described above, it is sometimes advocated to use yield spreads
instead of sovereign ratings as a proxy for default risk, since large yield spreads correspond to
high risk. Yield spreads refer to the difference between sovereign yields and US treasury bill
yields of the same maturity. Market yields are less stable, fluctuating daily and sometimes
substantially.

Figure 2: Precrisis and postcrisis yield spreads on an aggregate basis

Emerging non-
700 crisis countries
600 Crisis countries
Basis points

500
400
300
200
100
0
June 1997 June 1998

It appears that the use of market spreads rather than country ratings is not more efficient.
Indeed, for Asian countries, spreads have substantially widened after the crisis. As exhibited by
Figure 2, spreads were roughly of the same order of magnitude before the crisis. While spreads of
non-crisis countries have widened by less than 100% after the crisis, spreads of crisis countries
have more than tripled. Consequently, we conclude that spreads provide about the same
information as sovereign ratings do, and are much more volatile. This conclusion can be extended
to the Brazilian and the Russian crises. This discussion implies that yield spreads are
characterized by a lack of predictive power and cannot be used to obtain a reliable early warning
of country insolvency. This latter conclusion is confirmed by Mathieson and Schinasi (1999).

1.4 Recursive versus non-recursive models

The recent literature on country risk ratings contains several studies (Cantor and Packer, 1996,
Haque et al., 1998, Monfort and Mulder, 2000) which use multiple regression. The set of
independent variables used by Haque et al. (1998) as well as Monfort and Mulder (2000) includes
the lagged sovereign ratings of Standard & Poor, or Moody’s, or The Institutional Investor. The
correlation levels between the ratings of various agencies and the predicted values or ratings
obtained using multiple regression models referenced above are remarkably high.
To illustrate the actual meaning and importance of these results, we shall examine the
approach taken by Haque et al. (1998). That paper uses as its eight independent variables seven
macro-economic variables4, and the lagged rating, i.e. it includes The Institutional Investor
ratings both at times t and t-1, the former as a dependent variable, and the latter as an independent
one. It is important to note that country risk ratings are very stable, as shown by the transition
probabilities of the ratings published by Standard & Poor (1999) (see Table 19 in Appendix) and
Nickell et al. (2000).

4
T-bill rate, GDP growth rate, inflation rate, exports growth rate, ratio current account to GDP, the ratio of external
debt to GDP, the ratio of reserves to imports.
RRR 08-2004 PAGE 5

The 98% correlation level5 between The Institutional Investor ratings published respectively in
September 1997 and September 1998 confirms the stability property of sovereign ratings. In light
of this fact, the excellent correlation levels achieved by utilizing lagged ratings among the
independent variables can be attributed to a certain – possibly large – extent to this stability, and
may not necessarily give indications about the predictive power of the economic and political
variables used as predictors.
Although Cantor and Packer (1996) do not include the lagged ratings in their set of predictors,
they create a dummy variable, which is determined by the past ratings issued by Standard and
Poor (Claessens and Embrechts, 2002). This dummy variable is defined to be equal to 1 if a
country has ever been rated D or SD by Standard & Poor since 1970, and equal to 0 otherwise.
Even though their regression R-square is above 90%, their results are criticized by Claessens and
Embrechts (2002) and Jüttner and McCarthy (2000). Claessens and Embrechts mention that the
dates of the explanatory variables are not consistent, e.g. the values of some variables are
measured in 1994 or 1995, while that of others are averages for the period 1991-1994 or 1992-
1994. On the other hand, Jüttner and McCarthy evaluated the regression model of Cantor and
Packer for some other years, concluding that for 1998, it loses its predictive power. A recent
paper of Hu et al. (2002) develops a model using ordered probit to estimate country ratings. Their
model has an 83% correlation level and relies on economic variables and rating history of
countries.
A common feature of the econometric models above is the direct or indirect inclusion of
information derived from past Standard & Poor ratings (lagged ratings, rating history) among
their independent variables. A major drawback of such rating models is the impossibility of
applying them to not-yet-rated countries.

1.5 Objectives and main results

Our discussion in the preceding subsections indicates a need for making country risk ratings
more (i) transparent and (ii) consistent. A third criterion we would like to impose on an ideal
country risk rating system is that of (iii) self-containment, i.e. its non-reliance on any other past or
present country risk ratings. Clearly, this requirement precludes the use of lagged ratings as
independent variables. It is important to note that this approach is in marked contrast with that of
the current literature (discussed in the previous subsection), which does rely in one form or
another on lagged ratings. Finally, a fourth requirement imposed on the model is its (iv) stability,
i.e. extensibility both to subsequent years and to previously non-rated countries.
The wide acceptance of several of the major rating systems indicates that, while they may not
be perfect, they provide the currently best known evaluation of country risk. It is therefore
reasonable to base the design of any new rating system on one of the existing ones.
The central objective of this paper is to develop a transparent, consistent, self-contained, and
stable system, closely approximating the country risk ratings provided by a major rating agency.
We have selected the Standard & Poor country risk rating system as a benchmark for the desired
system. It is to be expected that, on the one hand, in most cases the ratings of the new system
should closely resemble those of Standard and Poor, and on the other hand, in the (hopefully few)
cases where the two ratings differ, the objective reasons, which determine the ratings of the
proposed model, should be justified by subsequent developments.

5
134 countries are considered.

5
PAGE 6 RRR 08-2004

In line with the existing literature, we use in the first part of this paper the technique of
multiple linear regression to achieve our objectives. We shall call the proposed system a non-
recursive multiple regression model of the S&P country rating system. In the second part of the
paper, we reanalyze the same problem, using this time a combinatorial-logical technique, in order
to derive a set of “logical rating scores” of countries, and show that they turn out to be
surprisingly similar to both the S&P ratings and the non-recursive regression scores.
The fact that the traditional linear regression model and the qualitatively different
combinatorial-logical model produce almost identical results strongly reinforces the confidence in
the results and in the validity of both models. In addition to the main result, we also demonstrate

• can be successfully applied to the rating of yet non-rated countries,

that the new rating systems

• are temporally stable,

• are consistent with the ratings of other agencies (Moody and The Institutional
Investor).

1.6 Paper structure

The paper is structured as follows. Section 1 describes the data considered and selected for use
in this paper. We provide a thorough literature review (see references in Tables 13 and 14) and
describe the selection of explanatory variables.
In Section 2, we use the 1998 Standard & Poor country risk ratings to develop a non-recursive
multiple regression model for the ratings considered as the dependent variable, regressed on a set
of economic and political variables (considered as the predictor variables). To evaluate the
accuracy of linear regression predictions, we use the k-folding cross-validation technique. We
show that the model correlates well not only with the ratings of Standard & Poor (95%), but also
with those of Moody’s (94.6%) and The Institutional Investor (93.6%).
We also evaluate the stability of the constructed non-recursive regression model in two ways.
First, we show the temporal stability of the non-recursive multiple regression model derived from
the 1998 dataset by applying it to the 1999 data. Second, we show that the proposed model can
successfully predict the ratings of several previously non-rated countries.
In Section 3, we analyze the same problem by using a combinatorial-logical technique, the
logical analysis of data (LAD) (Hammer, 1986, Crama et al., 1988, Boros et al., 2000), for
developing a model that evaluates the creditworthiness of countries. Based exclusively on the
S&P ordering of countries by their riskiness, i.e., without making any assumptions about the
magnitude of differences between consecutive ratings, we construct a discriminant function,
which provides an approximate measure of “relative riskiness” of an arbitrary country i compared
with another country j. The discriminant is expressed as a highly nonlinear polynomial in binary
variables, which indicate whether the values of relevant economic and political attributes of one
country do or do not exceed the values of the corresponding attributes of another country by
certain thresholds. The discriminant does not involve any information about past ratings. Further,
it is shown how to calculate a numerical measure of country risk (the “logical rating score”) in
such a way that the differences between the logical rating scores of all pairs of countries provide
the best L2-approximation of the corresponding discriminant values.
Subsection 4.1 introduces the concept of a pseudo-observation Pij, associated to a pair of
countries i, j, which provides a comparative description of i and j in the form of a multi-
dimensional vector, whose components are the differences of the values of those economic and
RRR 08-2004 PAGE 7

political attributes of countries i and j which were identified in the first part of the study. An
additional component of a pseudo-observation is an indicator which takes the value 1 (-1, 0) if the
country i in the pseudo-observation has a higher (lower, identical) rating than the country j.
The fundamental idea of this study is that a rating system can be essentially reconstructed
from the knowledge of the relative orderings of all pairs of rated countries. In other words, all that
matters in a rating is the qualitative order relation between countries, but not a quantitative
measure of the magnitude of differences between ratings.
The study focuses on deriving a model of the order relation between countries using the LAD
methodology which is briefly described in Subsection 4.2. Non-statistical and highly nonlinear,
this methodology is not restricted by the satisfaction of the assumptions underlying econometric

Based on the patterns of the LAD model, a discriminant, ∆(Pij), called relative preference, is
techniques.

computed for each pseudo-observation, Pij, The value of ∆(Pij) indicates whether country i should
be rated higher or lower than country j. Relying on the assumption that the ∆(Pij) values provide
good approximations of the differences of the ratings, the relative preferences are used to derive
an approximation of the ratings called logical rating scores of countries; these are calculated

set of countries, the rating of a country k is the regression coefficient, β k , in the regression model
using multiple linear regression, as described in Subsection 4.3. More precisely, denoting by I the

∆( Pij ) = βi − β j , for all i, j ∈ I , i ≠ j .

Subsection 4.4 is devoted to a thorough analysis of the results provided by the logical rating
score model. It is shown that the correlation between the logical rating scores and the S&P ratings
exceeds 95%. Moreover, the correlation between the logical rating scores and the scores provided
by the non-recursive linear regression model of Section 3 exceeds 98%.
The robustness of the model was confirmed by jackknife cross-validation. Furthermore, the
discriminant derived on the basis of the 1998 S&P ratings was applied to the 1999 data, and the
correlation between the resulting logical rating scores and the 1999 S&P ratings is shown to
exceed 94%. Since the discriminant does not involve any information about past ratings, it can be
used to derive the logical rating scores of previously non-rated countries; this is done in Section
4.4 where it is also shown that the logical rating scores are in agreement with subsequent S&P
ratings.
Section 5 presents general conclusions of this study.

2 Data

2.1 Sources
In this paper, we focus on the Standard & Poor country risk ratings. The risk of default is
generally defined by Standard & Poor as the probability that a sovereign obligor fails to meet a
principal or interest payment on the due date and in full. Standard & Poor’s ratings are based on
the information provided by the debtors themselves and by other sources considered reliable.
Standard & Poor provides sovereign ratings for local and foreign currency debt. In this paper,
we used the foreign currency sovereign ratings. Countries are more vulnerable to foreign
currency obligations. An obligor's capacity to repay foreign currency obligations may be lower
than its capacity to repay obligations in its local currency, owing to the sovereign government's

7
PAGE 8 RRR 08-2004

relatively lower capacity to repay external versus domestic debt. As noted by Cantor and Packer
(1996), foreign currency ratings remain the decisive factor in the international bond market.
Indeed, foreign currency obligations are more likely to be acquired by international investors than
domestic obligations. Foreign currency ratings reflect economic factors, as well as the country
intervention risk, i.e. the risk of a country imposing, for example, exchange controls or a debt
moratorium, while local currency ratings exclude country intervention risk.
Table 15 in the Appendix lists the different country risk levels or labels used by Standard &
Poor, and also provides descriptions associated with these labels. Countries which are assigned a
label inferior to BB+ are considered as non-investment grade (speculative) countries. Countries
rated CCC+ or lower are regarded as presenting serious default risks. BB indicates the least
degree of speculation and CC the highest. Ratings labeled from AA to CCC can be modified by
the addition of a plus or minus sign to show relative standing within the major rating categories.
We consider such subcategories as separate ratings in our analysis.
We have converted the Standard & Poor rating scale (ranging from AAA to SD) into a
numerical scale (ranging from 21 to 0) (see Appendix, Table 16) and shall liberally refer to both
of them as S&P ratings. This type of conversion is commonly used in the literature, see e.g.,
Bouchet et al. (2003), Estrella (2000), Ferri et al.(2001), Kräussl (2000), Monfort and Mulder
(2000), Mulder and Perelli (2001), Hu et al. (2002), Sy [2003]. Moreover, Bloomberg, a major
provider of financial data services, developed a standard cardinal scale for comparing Moody’s,
S&P and Fitch-BCA ratings (Kaminsky and Schmukler, 2002); in this scales, a higher numerical
value denotes a higher probability of default.
In Table 16 of the Appendix, we display Standard & Poor’s foreign currency sovereign ratings
of 69 countries published at the end of December 1998. Standard & Poor rates a limited number
of countries, with a special focus (at least in the past) on the industrial ones. However, in the last
decade, the number of Asian, Latin American and Eastern European economies rated by Standard
& Poor has significantly increased. We refer the reader to Hu et al. (2002) for the evolution of the
number of countries rated by Standard & Poor.
As mentioned above, country risk ratings encompass economic, financial and political aspects.
The statistical data of the economic and financial variables considered in this paper come from
the International Monetary Fund (World Economic Outlook database), from the World Bank
(World Development Indicators database) while those about the ratio of debt to gross domestic
product come from Moody’s publications. Values of political variables are provided by
Kaufmann et al. in two papers (1999a,b) that are joint products of the Macroeconomics and
Growth, Development Research Group and Governance, Regulation and Finance Institutes which
are affiliated with the World Bank. Before describing the relevance of the selected variables, we
discuss in Section 2.2 the selection method used.

2.2 Variable selection criteria

As underlined by Bilson et al. (2001), the selection of variables lends itself to criticism due to
the subjectivity and arbitrariness involved in this process. In this paper, the selection of relevant
variables is based on three criteria.
The first criterion is the significance of variables for estimating a country’s creditworthiness.
We have performed an extensive literature review which played an important role in defining the
set of candidate variables for inclusion in our model. Tables 13 and 14 list variables that have
been considered in the existing literature on country risk.
RRR 08-2004 PAGE 9

The second criterion is the availability of complete and reliable statistics. We want to avoid
difficulties related to missing data that could reduce the statistical significance and the scope of
our analysis. For instance, according to recent information received from The World Bank6, their
research concentrates on developing economies and they have data on the debt of 137 countries to
whom they loan funds and who report their external debt to The World Bank. Since high income
countries do not receive World Bank funds, they do not report their debt numbers to The World
Bank. Such situations have significantly complicated the process of compiling complete debt
statistics. Hu et al. (2002) also report the problem of data availability.
The third criterion is the uniformity of data across countries. We have considered, for
example, incorporating the unemployment rate statistics disclosed by the World Bank. However,
the World Bank underlines that unemployment is analyzed and compiled according to definitions
which differ from country to country.
It is worth noting that in addition to the variables listed in Tables 13 and 14 (see Appendix),
Haque et al. (1996), Cantor and Packer (1996), Larrain et al. (1997), Monfort and Mulder (2000)
and Hu et al. (2002) use a dummy variable that represents the historical solvency of a country.
Haque et al. (1996) use the lagged rating at time (t-1) as an independent variable in their
regression model. Monfort and Mulder (2000) claim that membership in the OECD is likely to be
a significant indicator for country risk ratings. The same authors emphasize also the importance
of the location of countries, by adding to their set of independent variables two dummy variables
to characterize the country’s location in Asia or in Latin America. Hu et al. (2002) also use
regional dummy variables.

2.3 Selected variables

Based on the criteria of relevance, availability and uniformity described above, we have
decided to incorporate the following variables7 in our model:
Gross domestic product per capita8 (GDPc): the GDP is the sum of gross value added by all
resident producers in the economy plus any product taxes and minus any subsidies not included in
the value of the products. It is calculated without making deductions for depreciation of
fabricated assets or for depletion and degradation of natural resources. The GDP is an indication
of the capacity of the government to solve a balance-of-payments crisis without having to default
on external debt. The larger the GDP, the wider the potential tax base and thus the higher the
ability of the government to fulfill its external obligations. The GDPc is a measure of the relative
wealth of a country and its level of development. The gross domestic product (GDP) is converted
to international dollars using purchasing power parity rates. The international dollar has the same
purchasing power over GDP as the U.S. dollar has in the United States.
Inflation rate (IR): the inflation rate is the change in the national price level between two periods.
The inflation rate used in our study is based on the consumer price index and is the annual
percentage change in the cost to the average consumer of acquiring a fixed basket of goods and
services. High inflation rates indicate structural problems in the country’s finances and may lead
to sovereign economic crises, as governments hike interest rates sharply in order to strengthen
their countries’ currencies. Should a country be unable or unwilling to pay the current budgetary

6
Anat Lewin, Private Communication, Development Data Group, The World Bank, June 06, 2001
7
Acronyms in parentheses following the name of variables are used in tables and appendices for referring to
variables.
8
Calculated on the basis of purchasing power parity in international dollars.

9
PAGE 10 RRR 08-2004

expenses, it must resort to inflationary money financing. High inflation rate results in a
substantial consumers’ purchasing power reduction and increases political discontent.
Trade balance (TB): trade balance is the balance of trade in goods expressed as a percentage of
GDP (purchasing power parity-PPP). This is the difference in value between a country's total
imports and exports (including information of oil and non oil exports, consumer goods, capital
goods) measured in current U.S. dollars divided by the value of GDP converted into international
dollars using purchasing power parity rates.
Exports’ growth rate (EGR): annual growth rate of exports of goods and services based on
constant local currency. Exports of goods and services represent the value of all goods and other
market services provided to the rest of the world. They include the value of merchandise, freight,
insurance, transport, travel, royalties, license fees, and other services, such as communication,
construction, financial, information, business, personal, and government services. They exclude
labor and property income as well as transfer payments. Countries having a high export growth
rate are expected to be more creditworthy. Indeed, exports are the primary source of foreign
currency inflows and therefore have a significant influence on the capacity of the country to
finance imports and service debt obligations.
International reserves (RES): this variable refers to gross international reserves, expressed in
terms of the number of months for which the existing reserves can cover the cost of imports of
goods and services. It gives an indication of the short-term capacity of an economy to meet its
imports obligations. The higher the value of RES, the lower the risk of default and the higher the
creditworthiness.
Fiscal balance (FB): fiscal balance is approximated by the ratio of central government financial
balance (surplus or deficit) to GDP. The central government’s balance represents the yearly fiscal
balance. Fiscal balances and debt stocks of governments are crucial indicators when analyzing
sovereign risk. The ability of governments to extract revenues from taxpayers and users of
services is a key factor that helps to determine whether governments will be able to make full and
timely payments of interest and principal on outstanding debt.
Debt to GDP (DGDP): here debt refers to the general government debt. The general government
debt as defined by the IMF (2001) includes “the consolidated budgets of the central,
state/regional, and local governments, along with the social security system and other extra-
budgetary funds engaged in noncommercial activities. Excluded are lending and refinancing and
the assets/liabilities of commercial state-owned or guaranteed enterprises, except for any net
financial transfers made as subsidies to these enterprises”. This balance, i.e., the difference
between total revenues and total expenditures, determines the net borrowing requirement of
general government, which can be met only by running down financial assets or borrowing net
new resources from the public and, thereby, adding to debt.
We have considered incorporating the unemployment rate and the ratio of the current account
balance to GDP. While the latter turned out to be redundant with trade as a percentage of GDP,
the former has been excluded from consideration due to the lack of consistency in its definition.
As noted by the World Bank, the treatment reserved to temporarily laid off workers, to those
looking for their first job, and the criteria referred to for being considered as unemployed, differ
significantly between countries.
For political variables, it is very difficult to find reliable and complete data. In our model, we
have considered the six variables provided by Kaufmann et al. (1999a). These six variables are:
political stability and violence, voice and accountability, government effectiveness, regulatory
burden, corruption , rule of law. These variables are viewed as capturing the fundamentals of the
RRR 08-2004 PAGE 11

governance concept defined as “the traditions and institutions by which authority in a country is
exercised” (Kaufmann et al.,1999a).
As emphasized by Kaufmann et al. (1999, a and b), political stability and voice and
accountability both refer to the process by which governments are elected, monitored and
replaced. Government effectiveness and regulatory burden reflect the capacity of the government
to adopt sound policies. Corruption and rule of law are proxies for the “respect of citizens and
institutions for the rules which govern their interactions”. In order to avoid or at least limit
redundancies in our model, we select only one variable for each dimension of governance. We
have selected:
Political stability (PS),
Government effectiveness (GE), and
Corruption (COR).
The higher the values of these variables, the less likely the country is to default9. The variables
are defined on a (-3.5, 3.5) interval and are based on estimations provided by polls of experts and
cross-country surveys.
The variables we have described so far have been considered previously in the literature and
are available in the form used in our study (as ratios or as growth rates). We have also decided to
construct a new variable (ER) and to add a variable (financial depth and efficiency) which, to the
best of our knowledge, has not been used before in country rating studies. Here are the
descriptions of these two variables:
Exchange rate (ER): is defined as the ratio of the current value of the exchange rate to the
moving average of the real effective exchange rate10 over five years (1994 to 1998). While the
exchange rate has been used in previous country rating studies, we consider the ratio introduced
here to be more significant, since it indicates the dynamics of changes in the exchange rate, by
specifying whether the trend is up (ER>1) or down (ER<1).
Financial depth and efficiency (FDE): is represented by the ratio of the domestic credit provided
by the banking sector to the GDP. Households accumulate claims on financial institutions that,
acting as intermediaries, pass funds to final users. Correlated to the development of the economy,
the indirect lending by savers to investors becomes more efficient and gradually increases assets
relative to the GDP. Viewed from this perspective, the ratio of domestic credit to the GDP
reflects the financial depth and efficiency of the country’s financial system. More specifically,
this variable is used to measure the growth of the banking system since it reflects the extent to
which savings are financial. To our knowledge, the financial depth and efficiency variable has not
been considered previously in the evaluation of country risk ratings.

2.4 Dataset content

In summary, on the basis of the considerations described above, we have constructed a dataset

• gross domestic product per capita, inflation rate, trade balance, international
involving nine economic/financial variables:

reserves, fiscal balance, exports growth rate, debt to GDP, financial depth and

9
The higher the value of the corruption variable, the less corrupted the considered country is perceived to be. This
variable can therefore be called “corruption quality”.
10
Real effective exchange rate is the nominal effective exchange rate (a measure of the value of a currency against a
weighted average of several foreign currencies) divided by a price deflator or index of costs.

11
PAGE 12 RRR 08-2004

efficiency, and exchange rate (we have used the values taken by these variables at
the end of 1998);

• political stability, government effectiveness and corruption level.

and three political variables:

We have compiled the values of these twelve variables for the sixty-nine countries considered:
24 industrialized countries, 11 Eastern European countries, 8 Asian countries, 10 Middle Eastern
countries, 15 Latin American countries and South Africa. We use the Standard & Poor country
risk ratings for these countries at the end of December of 1998.

3 A Statistical Model

3.1 Non-recursive multiple regression model

3.1.1 The model, results, and consistency with S&P

In order to derive a non-recursive model of Standard & Poor’s ratings, we shall fit the

Y = α + ∑ βi * X i + ε ,
regression equation:
M
(1.1)
i =1
where the dependent variable Y is the country risk rating given by Standard & Poor at the end of
December 1998 (or more precisely a numerical representation of Standard & Poor’s ratings), the
independent variables X i are the economic and political variables described in Section 2.1, and
ε is the error term. In view of the desired non-recursiveness of the model, the independent
variables do not include directly or indirectly ratings of previous years. Results given in this
section have been obtained using the SPSS statistical package.
The proposed model exhibits an excellent fit, with the coefficient of multiple determination R-
square being 91.2%, and the adjusted R-square 89.3%. The multiple correlation level between the
observed values (i.e. the Standard & Poor ratings) and the predicted ones (i.e. the ratings given by
the non-recursive regression model) is equal to 95.5%. The later are given in Appendix (Table 17,
Column 2).
Table 1 below details how the regression equation accounts for the variability in the response
variable, the last column giving the statistical level (1-p) at which the model is significant.

Table 1: Analysis of variance (ANOVA)

Sum of Squares Degrees of Mean Square F-statistic p-value
freedom
Regression 1609.546 12 134.129 48.458 0.000
Residual 155.005 56 2.768
Total 1764.551 68
RRR 08-2004 PAGE 13

Table 2 presents the regular and the standardized regression coefficients (i.e., those
corresponding to the model fitted to standardized data). The last column in Table 2 indicates
whether the corresponding independent variable is statistically significant (at the confidence level
of 1-p).
Table 2: Regression results
Variables Unstandardized Standard error Standardized t-statistic p-value
coefficients coefficients
(Beta)
Intercept 8.769 .860 10.195 0.000
FDE 1.693E-02 0.007 0.148 2.513 0.015
RES 0.116 0.101 0.055 1.148 0.256
IR -2.831E-02 0.018 -0.080 -1.557 0.125
TB -1.192E-02 0.006 -0.094 -1.960 0.055
EGR -1,218E-02 0.031 -0.017 -0.396 0.694
GDPC 3.081E-04 0.000 0.499 5.294 0.000
ER -1.968E-02 0.011 -0.079 -1.768 0.083
FB 0.120 0.086 0.078 1.393 0.169
DGDP -1.610 0.795 -0.091 -2.026 0.047
PS 1.378 0.533 0,197 2.584 0.012
GEF 1.977 0.920 0.316 2.149 0.036
COR -0.605 0.842 -0.111 -0.718 0.476
At the 5% significance level, it appears that five independent variables are statistically
significant. These are: financial and depth efficiency (FDE), gross domestic product per capita
(GDPc), ratio debt to gross domestic product (DGDP), political stability (PS) and government
efficiency (GE).
The regression results described above indicate that the non-recursive regression model has an
excellent fit with the data. However, the excellence of the fit does not automatically guarantee the
predictive power of the model, if the model violates some of the critical assumptions of multiple
regression theory, as is the case with proposed model. Indeed,
• there is a strong correlation between some of the variables considered, e.g. between
the political variables, especially government efficiency and corruption, possibly
leading to difficulties related to multicollinearity, and the ill-conditioned nature of
the resulting matrix;
• the predictors are not normally distributed;
• if too many predictor variables are used relatively to the number of observations,
fitting multiple regression can lead to overfitting, and the estimates of the regression
line can be unstable and the results may not be reproducible; the number of
variables used is generally recommended to be no more than 5 to 10% of the
number of observations, which is clearly not the case of this study that involves 69
observations and 12 variables.
In view of these issues, it is surprising that the cross-validation results presented in the next
section provide a strong confirmation of the predictive power of the non-recursive regression
model presented above.

13
PAGE 14 RRR 08-2004

3.1.2 Cross-validation

To validate the predictive power of the non-recursive regression model, we use a resampling
technique known as cross-validation, and more specifically, a popular variant of it called k-
folding (e.g., Shao, 1993, Shao and Tu, 1995, Efron, 1982, Hurvich and Tsai, 1989, Hjorth 1994,
Breiman and Spector, 1992). In k-folding, observations are divided into k subsets of
approximately equal size. The regression model is trained k times, each time leaving out from
training one of the k subsets, and using the omitted subset to test the regression-predicted country
risk rating. In this paper, based on the relatively small size of the sample, we have selected k to be
10, and partitioned the sample into 10 groups of 6 or 7 countries each. The groups were selected
using stratified random sampling, i.e. assuring that each group contains about the same number of
investment-, speculative- and default-grade countries (see S&P’s classification, Table 15).
In Appendix (Table 17, Column 3) we present the in-the-sample predictions of the non-
recursive multiple regression model obtained in the preceding Section, and the out-of-the-sample
predictions obtained using the 10-fold cross-validation. The major results are the following:
• the correlation between the in-the sample and the out-of-the sample predictions is

• the correlation between the Standard and Poor ratings and the out-of-the sample
99.1%,

predictions is 95.6%.
The very high correlation levels demonstrate clearly that the impressive results of Section
3.1.1 are not due to chance or overfitting.

3.1.3 Rating discrepancies between S&P and the proposed model

In this section, we shall identify the few countries for which the predictions of the non-
recursive regression model disagree with the Standard & Poor ratings. In order to accomplish
this, we shall construct confidence intervals for our predicted ratings11.
Let us introduce some notations. Let n and p refer to the number of observations and
predictors, respectively. The expression t (1 − α / 2, n − p ) refers to the Student test with
( n − p ) degrees of freedom, and with upper and lower tail areas of α / 2 . Let X j be the p-
dimensional vector of the values taken by the observation Y j on the p predictors, while X 'p be the
transposed of X j . Let the expression ( X ' X ) −1 refer to the variance-covariance matrix, i.e. the
inverse of the [ p × p ] -dimensional matrix ( X ' X ) . Denoting by MSE the mean square of errors in
^
the regression, the estimated variance s 2 [Y j ] of the predicted rating is:

s 2 [Y j ] = MSE *[ X 'j ( X ' X ) −1 X j ] ,

^
(1.2)

while the (1 − α ) -confidence interval for the predicted rating Y j is:

^

{Y j − t (1 − α / 2, n − p ) * s[Y j ], Y j + t (1 − α / 2, n − p ) * s[Y j ]}
^ ^ ^ ^
(1.3)

11
All formulae given in this section as well as those in Section 3.2.3 are from Neter et al. (1996)
RRR 08-2004 PAGE 15

We say that there is a discrepancy between the Standard & Poor rating R SPj of a country j and

ours, if the Standard & Poor rating is not in the confidence interval, i.e.:
j ∉ {Y j − t (1 − α / 2, n − p ) * s[Y j ], Y j + t (1 − α / 2, n − p ) * s[Y j ]} for α = 0.1
^ ^ ^ ^
R SP (1.4)
Taking α equal 5%, this formula identifies four discrepancies. Three countries (Iceland,
Pakistan and Argentina) are rated higher by the non-recursive regression model than by Standard
& Poor, while Columbia is rated higher by Standard & Poor. It is remarkable that subsequently
the Standard & Poor ratings for two of these four countries (Columbia and Pakistan) have been
modified in the direction suggested by the regression model. More precisely, Columbia has been
downgraded by Standard & Poor twice, moving from BBB- in December 1998 to BB+ in
September 1999, and then to BB in March 2000. After being downgraded in January 1999 (SD),
Pakistan was upgraded to B- in December 1999. On the other side, Iceland’s rating has remained
unchanged, and Argentina’s rating has endured significant downgrade, which started however
only in November 2000.

3.1.4 Are political variables necessary?

In this section, we test the predictive power of the non-recursive regression model, from which
the three political variables are omitted. The R-square as well as the adjusted R-square of this
model are equal to 88.6 % and 86.9 % respectively. These values are lower than the
corresponding values for the original non-recursive regression model, indicating a loss in
predictive power resulting from the omission of the three political variables. The predicted ratings
are given in Appendix (Table 17, Column 4).

Table 3: Regression coefficients

Variables Unstandardized Standard error Standardized t-statistic p-value
coefficients coefficients
Intercept 1.463 2.631 0.556 0.58
FDE 2.20E-02 0.007 0.192 3.058 0.003
RES 0.194 0.107 0.091 1.818 0.074
IR 1.22E-02 0.034 0.034 0.359 0.721
TB -5.85E-03 0.006 -0.046 -0.901 0.371
EGR 2.75E-02 0.032 0.039 0.849 0.399
GDPC 4.38E-04 0 0.709 10.776 0
ER 6.933 2.737 0.23 2.533 0.014
FB 0.244 0.083 0.158 2.933 0.005
DGDP -1.815 0.853 -0.102 -2.129 0.037

It appears that five of the independent variables are statistically significant at a 95% level.
These variables are financial and depth efficiency (FDE), gross domestic product per capita
(GDPc), debt to gross domestic product ratio (DGDP), exchange rate (ER) and fiscal balance
(FB). The correlation coefficient between the predicted ratings and those of Standard & Poor is
equal to 94.14 % and is lower than in the original model. Moreover, the inferior fit of this model
results in wider confidence intervals as compared to the original model.

15
PAGE 16 RRR 08-2004

The discrepancies between Standard & Poor’s predictions and ours involve four countries
(Russia, Pakistan, South Korea and Iceland), all being underrated by Standard & Poor. The
ratings of three of these countries (Russia, Pakistan, South Korea) have been modified since, in
the direction suggested by our model, while the rating of Iceland has remained unchanged. The
evolution of ratings for Pakistan has already been described in Section 3.1.3. South Korea has
both been upgraded three times, moving from BB+ in December 1998 to BBB+ in November
2001. Russia has first been downgraded to SD in January 1999, before being upgraded three
times and being rated B+ in December 2001.
In conclusion, the model which omits political variables appears to be somewhat less closely
related to the S&P model which it is supposed to reflect, but on the other end this apparent
weakening is not sufficiently clear to allow us to draw any definite conclusions. It should be
added here that the economic variables are easier to obtain than the political ones, which are
published less frequently.

3.2 Stability of the non-recursive regression model

3.2.1 Consistency with ratings of other agencies

In addition to analyzing the correlation level between Standard & Poor’s ratings and those of
the proposed non-recursive model, the latter has to be compared with the ratings of other
agencies, e.g. Moody’s and The Institutional Investor. We present below the results of these
comparisons, based on Moody’s and The Institutional Investor ratings issued at the end of
December 1998 and in March 1999 respectively. We shall start by presenting a brief description
of the rating systems of Moody’s and of The Institutional Investor.
Moody’s sovereign ratings are defined, as “a measure of the ability and willingness of the
country’s central bank to make available foreign currency to service debt, including that of
central government itself” (Moody’s, 1995). Similarly to Standard & Poor, Moody’s uses a
nominal rating scale (Table 18 in Appendix), which contains the same number of categories as
Standard & Poor’s ratings. A large proportion of countries receive the same rating from Moody’s
and Standard & Poor, and when they are different, the difference is usually not more than one
notch.
The Institutional Investor country risk ratings were first compiled in 1979, and are published
now regularly, in March and September of every year, for an increasing number of countries,
which reached 145 in 2000. The Institutional Investor ratings are numerical, ranging from 0 to
100, with 100 corresponding to the lowest chance of default. The Institutional Investor relies on
evaluations of the creditworthiness of the countries to be rated, provided by economists and
international banks, each respondent using their own criteria. Responses are aggregated by The
Institutional Investor, greater weights being given to responses from institutions with higher
worldwide exposure.
The correlation levels between the ratings given by the non-recursive multiple regression
model and those given by Standard and Poor, Moody’s and The Institutional Investor are reported
in Table 4.
RRR 08-2004 PAGE 17

Table 4: Correlation between the non-recursive model and other ratings

Non-recursive The Institutional
Standard & Poor Moody’s
regression Investor
Non-recursive
100% 95.05% 94.56 % 93.60 %
regression
Standard & Poor 100% 98.01% 96.18%

Moody’s 100% 96.31%

The Institutional
100%
Investor
It can be seen that the very high correlation levels between the ratings given by the non-
recursive regression model and those given by the three rating agencies underline the relevance of
the proposed model.

3.2.2 Temporal stability of the non-recursive regression model

3.2.2.1 1999 data

In order to supplement the indications of stability of the non-recursive regression model
provided by cross-validation, we shall test its temporal stability by extending the analysis to the
data of the following year (1999). Using as input the same economic and political variables
described in Section 2.3, we shall use the 1999 data in two ways. First, we shall use the model
derived from the 1998 data to check the quality of its predictions when applied to the 1999 data.
Second, we shall derive a new model based on the 1999 data and check the goodness of its fit.
In this experiment, we have used 1999 data with two exceptions. First, in the 1999 data, 16 out
of the 828 variable values (1.9%) are missing; for these missing values, we have substituted their
corresponding values taken in the previous year. Second, the political variables, reflecting the
perceptions of governance quality by a large number of survey respondents in industrial and
developing countries, as well as in non-governmental organizations, are not necessarily compiled
and updated on a yearly basis. The indices of Kaufmann et al. (1999a, 1999b and 2002) have
been published twice, referring respectively to data of 1998 and 2000-2001, but were not
compiled for 1999. The high degree of stability of political variable indices is reflected in the fact
that the correlation between the 1998 and the 2000-2001 indices varies between 95% and 98%,
depending on the variable. In our calculations, we have approximated the values of the three
political variables appearing in our model (corruption, government efficiency and political
stability) for 1999, by averaging their values for 1998 and 2000-2001.

3.2.2.2 Applying the 1998 model to the 1999 data

The purpose of this section is to test the applicability of the non-recursive regression model
built on the 1998 data, for predicting the 1999 country risk ratings. More precisely, we substitute
the 1999 data into the regression model built on the 1998 data, and compare the results obtained
in this way with Standard & Poor’s 1999 ratings. The new predicted country risk ratings are given
in Appendix (Table 17, Column 5).
The most important result of this experiment is the very high level of correlation (94.74%)
between the predicted ratings and the 1999 Standard & Poor ratings, confirming the consistency
and the temporal stability of the non-recursive regression model.

17
PAGE 18 RRR 08-2004

In order to identify the discrepancies, we have to recalculate the prediction confidence

intervals for the new, 1999 observations. Since these observations have not been used in deriving
the regression coefficients, formulae (1.2), (1.3), and (1.4) can no longer be used for constructing
the confidence intervals. The variance s 2 [ pred ] should now be computed as follows:
s 2 [ pred ] = MSE *[1 + X 'j ( X ' X ) −1 X j ] , (1.5)

while the (1 − α ) confidence interval for Y j ,n will be given by:

^

{Y j ,n − t (1 − α / 2, n − p ) * s[ pred ], Y j ,n + t (1 − α / 2, n − p ) * s[ pred ]}
^ ^
(1.6)
SP
We say that there is a discrepancy between the Standard & Poor rating R j and the non-
recursive regression model, if :
j ∉ {Y j ,n − t (1 − α / 2, n − p ) * s[ pred ], Y j ,n + t (1 − α / 2, n − p ) * s[ pred ]} for α = 0.1
^ ^
R SP (1.7)
The results show that the only three discrepancies between our ratings and those of Standard &
Poor concern Argentina, Iceland and Russia, all of these countries being underrated by Standard
& Poor. The cases of Argentina and Iceland have already been discussed in Section 3.1.3. As far
as Russia is concerned, it was first downgraded to SD in January 1999, but upgraded afterwards
to B- in December 2000, and to B in June 2001, thus confirming our prediction.

3.2.2.3 New model for 1999

To obtain a new non-recursive regression model for 1999, we proceed in the same way as for
1998, using as input the political and economic variables described above and the sovereign risk
ratings published by Standard & Poor at the end of December 1999. As a result, we obtain new
regression coefficients that constitute the 1999 model. The 10-folding cross-validation tests
performed showed a correlation level of 98.63% between the in-the-sample and out-of-the-
sample predicted ratings, thus validating the 1999 model.
The final non-recursive regression model for 1999 is given in Table 5 and the predicted ratings
are given in Appendix (Table 17, Column 6).
It is important to emphasize that the correlation between the 1999 Standard & Poor ratings and
those predicted by the 1999 non-recursive regression model is 96.4% (even exceeding that of
1998). The only discrepancies between the two models, determined using (1.4), concern Iceland
and Argentina, which appear to be underrated by Standard & Poor.
Table 5: Regression results
Variables Unstandardized Standard error Standardized t-statistic p-value
coefficients coefficients
Intercept 9.97 2.065 4.829 0
Intercept 9.97 2.065 4.829 0
FDE 0.007446 0.006 0.065 1.312 0.195
FDE 0.007446 0.006 0.065 1.312 0.195
RES -0.002063 0.093 -0.001 -0.022 0.982
RES -0.002063 0.093 -0.001 -0.022 0.982
IR -0.0875 0.025 -0.236 -3.454 0.001
IR -0.0875 0.025 -0.236 -3.454 0.001
TB -0.005184 0.005 -0.046 -1.003 0.32
TB -0.005184 0.005 -0.046 -1.003 0.32
EGR -0.023 0.023 -0.038 -0.989 0.327
EGR -0.023 0.023 -0.038 -0.989 0.327
GDPC 0.0002413 0 0.406 4.894 0
GDPC 0.0002413 0 0.406 4.894 0
RRR 08-2004 PAGE 19

ER -0.798 2.274 -0.025 -0.351 0.727

FB -0.05838 0.071 -0.041 -0.819 0.416
FB -0.05838 0.071 -0.041 -0.819 0.416
DGDP -0.856 0.727 -0.048 -1.178 0.244
DGDP -0.856 0.727 -0.048 -1.178 0.244
PS 0.786 0.569 0.108 1.382 0.173
PS 0.786 0.569 0.108 1.382 0.173
GEF 2.722 0.946 0.422 2.878 0.006
GEF 2.722 0.946 0.422 2.878 0.006
COR -0.04607 0.792 -0.008 -0.058 0.954
COR -0.04607 0.792 -0.008 -0.058 0.954

3.2.3 Rating previously non-rated countries

In this section, we test the prediction power of our model on a set of countries which were not
used for constructing our model. We have obtained data for four such countries (Ecuador,
Guatemala, Jamaica and Papua New Guinea), the ratings of which by Standard & Poor started
after December 1998. The ratings of these countries using the 1998 non-recursive regression
model (described in Section 3) with 1998 and 1999 data are presented in Table 6.

Table 6: Predicted ratings for previously non-rated countries

Time of the 1998 regression model rating

S&P linear using:
first S&P S&P rating
extension
rating 1998 data 1999 data
Ecuador 07/2000 (SD) (0) 6.03 5.69
Guatemala 10/2001 (BB) (10) 8.57 8.14
Jamaica 11/1999 (B) (7) 6.00 7.16
Papua New Guinea 01/1999 (B+) (8) 6.93 6.88

Comparing the ratings predicted by the non-recursive regression model with those given by
Standard & Poor, it can be seen that the model has a significant predictive power, even when
applied to countries not used in constructing the model. Indeed, three of the four countries above
(Papua New Guinea, Jamaica and Guatemala) have their first Standard & Poor ratings within the
95% confidence intervals of the predicted ratings. The rating of Ecuador is even more interesting.
While the first Standard & Poor rating of that country (SD) is not in the 95% confidence interval
of the predicted value, that rating (given in July 2000) was revised after only one month (in
August 2000) to B, which falls within the 95% confidence interval of our prediction.

4 A Combinatorial Model
4.1 Pairwise country comparisons: Pseudo-observations
Let us associate to every country i ∈ I = {1,…,69} considered in this study, the 13-
dimensional vector Ci, whose first component is the country risk rating given by Standard and
Poor, while the remaining 12 components specify the values of the nine economic/financial and
of the three political variables.

19
PAGE 20 RRR 08-2004

pairs of countries. For this purpose, we shall construct for every pair of countries i, j ∈ I, a
In this study, instead of considering countries independently of each other, we shall consider

pseudo-observation Pij , which shall provide in a way specified below a comparative description
of the two countries.
The pseudo-observations are represented as 13-dimensional vectors. The first component is an
indicator which takes the value 1 if the country i in the pseudo-observation Pij has a higher rating
(i.e., lower risk) than the country j, takes the value –1 if the country j has a higher rating than the
country i, and takes the value 0 if both countries have the same rating. The other components k , k
= 2,…,13 of the pseudo-observation Pij[k] are obtained simply by taking the differences of the

Pij [k ] = Ci [k ] − Ci [k ], k = 2,...,13
corresponding components:
(1.8)
One of the advantages of this transformation is that it allows us to avoid the problems posed
by the fact that the original dataset contains only a small number (| I |) of observations. The
transformation (1.8) provides a substantially larger dataset, which contains | I |*(| I | -- 1) pseudo-
observations. It will be seen that the potential problems created by the non-independence of
pseudo-observations can be overcome by robust data analysis techniques.
The fundamental idea of this study is that a rating system can be essentially reconstructed from
the knowledge of the relative standings of all pairs of rated countries. In other words, all that
matters in a rating is the order relation between countries. Therefore, this study will focus on
inducing a model for the order relation between countries.
In order to illustrate the construction of pseudo-observations, let us consider as an example the
case of Japan and Canada.
Table 7 reports the values taken by the twelve economic/financial and political variables, as
well as the rating given by Standard & Poor to these countries at the end of December 1998.
Table 7: Examples of country observations

S&P
FDE RES IR TB EGR GDPc ER FB DGDP PS GE COR
RATING
CJapan AAA 138.44 5.168 .65 21.7471 -2.54 24314.2 0.839 -7.7 0.47 1.153 0.839 0.724

CCanada AA+ 94.69 1.01964 0.99 55.9177 8.79 24855.7 0.939 0.9 0.5 1.027 1.717 2.055

In Table 8 below, we construct the pseudo-observation PJapan,Canada from the country

observations CJapan and CCanada given in Table 7. Since Sweden and Australia are rated
respectively AA+ and AA by Standard & Poor at the end of December 1998 and the rating AA+
is superior to the rating AA, the first component of the pseudo-observation vector takes the
value 1. Clearly the pseudo-observation PCanada,Japan is anti-symmetric to PJapan,Canada , as shown
in Table 8.
Table 8: Examples of pseudo-observations
Indicator FDE RES IR TB EGR GDPc ER FB DGDP PS GE COR
PJapan,Canada 1 43.75 4.15 -0.34 -34.17 -11.33 -541.5 -0.1 -8.6 -0.03 0.126 -0.878 -1.331
PCanada,Japan -1 -43.75 -4.15 0.34 34.17 11.33 541.5 0.1 8.6 0.03 -0.126 0.878 1.331
RRR 08-2004 PAGE 21

4.2 Logical Analysis of Data (LAD) – An Overview

The logical analysis of data is a combinatorics-, optimization-, and Boolean logic-based
methodology for analyzing archives of observations (Boros et al., 1997). Initially created for the
classification of binary data (Hammer 1986), LAD was later extended (Boros et al., 1997) to
datasets having binary (positive and negative) observations, depending on numerical variables.
LAD distinguishes itself from other pattern recognition methods and data mining algorithms by
its capacity to discover a minimal (irredundant) set of variables along with a collection of patterns
built on them, in order to explain the positive or negative nature of the observations in an dataset
(Boros et al., 2000).
Each observation is represented by an (n+1)-dimensional vector, the first component of which
is the classification of the observation, the n other variables being the inputs. The classification is
binary (0 or 1), i.e., it specifies the positive (1) or negative (0) nature of the observation.
The purpose of LAD is to discover a “function” f depending on the n input variables, or an
approximation of f , allowing the correct discrimination between positive and negative
observations. The approximation of f is constructed as a weighed sum of patterns. Positive
(negative) patterns are combinatorial rules which impose upper and lower bounds on the values

• a sufficiently high proportion of the positive (negative) observations in the dataset

of a subset of input variables, such that:

• a sufficiently high proportion of the negative (positive) observations violate at least one
satisfy the conditions imposed by the pattern, and

of the conditions of the pattern.

The conditions defining a pattern specify that the values of some of the variables are “large” or
are “small”; more precisely, these conditions require these values to be above or below certain
specified levels, called cutpoints. By associating an indicator variable to each cutpoint, the dataset
is “binarized”, i.e., each original numerical variable is replaced by several binary ones.
The following terminology will be useful. The degree of a pattern is the number of variables
the values of which are bounded in the definition of the pattern. The prevalence of a positive
(negative) pattern is the proportion of positive (negative) observations covered by it. The
homogeneity of a positive (negative) pattern is the proportion of positive (negative) observations
among those covered by it. Patterns of low degree, high prevalence and high homogeneity have
been shown to be the most efficient in LAD applications, e.g., (Boros et al., 2000).
The first step in applying LAD to the dataset is to generate the pandect, i.e., the collection of
all patterns in an dataset. The number of patterns contained in the pandect of a dataset of such
dimensions can be exponentially large, in the order of hundreds of thousands, possibly millions.
Because of the enormous redundancy in this set, we shall impose a number of limitations on the
set of patterns to be generated, by restricting their degrees (to low values), their prevalences (to
high values), and their homogeneities (to high values); these bounds are known as LAD control
parameters. It should be added that the quality of patterns satisfying these conditions is usually
much higher than that of patterns having high degrees, or low prevalences, or low homogeneities.
Several algorithms have been developed for the efficient generation of substantial subsets of the
pandect corresponding to reasonable values of the control parameters (Alexe and Hammer 2001;
Hammer, Kogan et al. 2001; Alexe, Alexe et al. 2002; Alexe and Hammer 2002).
The substantial redundancy among the patterns of the pandect makes necessary the extraction
of (usually small) subsets of positive and negative patterns, sufficient for classifying the
observations in the dataset. Such collections of positive and negative patterns are called models.
A model is supposed to contain positive (negative) patterns « covering » (i.e., whose conditions

21
PAGE 22 RRR 08-2004

are satisfied by) each of the positive (negative) observations in the dataset. Furthermore, good
models tend to minimize the number of points in the dataset covered by both positive and
negative patterns in the model.
The way a LAD model can be used for classification is the following. An observation (whether
it is contained or not in the given dataset) which satisfies the conditions of some of the positive
(negative) patterns in the model, but which does not satisfy the conditions of any of the negative
(positive) patterns in the model, is classified as positive (negative). An observation satisfying
both positive and negative patterns in the model is classified with the help of a discriminant
which assigns specific weights to the patterns in the model (Boros et al., 2000). More precisely, if
p and q represent the number of positive and negative patterns in a model, and if h and k represent

θ , then the value of the discriminant ∆ (θ ) is simply

the numbers of positive, respectively negative patterns in the model covering a new observation

∆ (θ ) = h / p - k / q, (1.9)

observation for which ∆ (θ ) = 0 is left unclassified, since the model either does not provide enough
and the corresponding classification is determined by the sign of this expression. Finally, an

evidence, or provides conflicting evidence; fortunately it has been seen in all the real-life
problems considered that the number of unclassified observations is extremely small.

4.3 Logical rating score (LRS)

In order to “learn” the Standard & Poor rating system, we shall proceed in two steps. In the
first step, the proposed model is derived only from the information indicating for each pair of

we shall derive an LAD model, whose discriminant provides a numerical measure ∆(Pij) of the
countries having different S&P ratings which of the two is rated higher. From this information,

“superiority” of the country i‘s rating over that of country j.

In the second step, applying multiple linear regression we derive new numerical ratings of all

way that their pairwise differences provide the best approximation of the numerical measures ∆
countries, called “logical rating scores” (LRS); the logical rating scores are obtained in such a

obtained in the first step. These scores will be shown in Section 5 to have a very high correlation
with the S&P ratings.

4.3.1 From pseudo-observations to relative preferences

The “observations” of the dataset used in the first step are those pseudo-observations Pij,
which correspond to countries i and j having different ratings. Each pseudo-observation Pij is
classified as positive or negative, according to the value of the indicator variable, i.e., depending
on whether i is rated higher than j or vice versa. Clearly, the training set is anti-symmetric.

∆ ( Pij ) for each pseudo-observation Pij (i ≠ j). The values ∆ ( Pij ) of the discriminant are called the
After having constructed the LAD model, we compute (according to (1.9)) the discriminant

relative preferences, and the [69 x 69]-dimensional anti-symmetric matrix, ∆, having them as
components will be called the relative preference matrix. While the LAD model was derived

matrix components are the values ∆ ( Pij ) (i ≠ j) taken by the discriminant for every pair of
using only those pseudo-observation Pij for which i and j were rated differently, the discriminant

countries, including those that have the same S&P ratings.

RRR 08-2004 PAGE 23

To illustrate the concept of the relative preference matrix, let us consider the example of three
countries, Japan, Canada, and Belgium, and the six associated pseudo-observations. The derived
relative preferences obtained from the LAD model are shown in Table 9.

Table 9: Relative preferences for Japan, Canada and Belgium

Japan Canada Belgium

Japan 0.00625 -0.00625
Canada -0.00625 0.03125
Belgium 0.00625 -0.03125
A naïve approach to the use of the relative preferences for deriving country ratings would

specifically, a large positive value ∆( Pi , j ) could be interpreted as country i being more

rely on the direct interpretation of their sign as an indicator of rating superiority. More

value of the relative preference ∆( Pi , j ) . A value equal or nearly equal to 0 would mean that the
creditworthy than country j , while the opposite conclusion could be drawn from a large negative

evidence for drawing conclusion about the relative creditworthiness of countries i and j is either
lacking or conflicting.
The difficulty of this naïve approach is that it overlooks the imprecision of (i.e., the noise
inherent in) data, and therefore of the relative preferences. The matrix above illustrates this
phenomenon, since the naïve interpretation would rate Japan above Canada, Canada above
Belgium, and at the same type Belgium above Japan, which contradicts the basic requirement of
transitivity of an order relation.
In the following section in order to overcome this difficulty, we shall relax the overly
constrained search for (possibly non-existent) country ratings whose pairwise orderings are in
precise agreement with the signs of relative preferences, to the more flexible search for logical
rating scores (LRS), having numerical values whose pairwise differences approximate well the
relative preferences.

4.3.2 From relative preferences to logical rating scores using regression analysis

It has been common practice in the research literature (see e.g., Ferri and Liu, 1999, Hu et al.,
2002, Kräussl, 2000, Monfort and Mulder, 2000, Sy, 2003) to interpret sovereign ratings as
cardinal values. Assuming that the sovereign ratings β can be interpreted as cardinal values, it is
natural to view the relative preferences ∆ as differences of the corresponding ratings:
∆( Pij ) = β i − β j , for all i, j ∈ I , i ≠ j (1.10)

∆( Pij ) = β i − β j + ε ij , for all i, j ∈ I , i ≠ j

Obviously, the system (1.6) may or may not be consistent. We shall therefore replace it by:
(1.10)
The determination of those values of the β’s which provide the best L2 approximation of the

∆(π ) = ∑ β k * xk (π ) +ε (π ) ,
∆’s can be found as a solution of the following multiple linear regression problem:
(1.11)
k∈I

π = {(i, j ) i, j ∈ I , i ≠ j}
where
(1.12)

23
PAGE 24 RRR 08-2004

1, for k = i
and


xk (i, j ) =  −1, for k = j
0, otherwise
(1.13)

The logical rating scores β k estimated by the regression model are presented in column 4 of
Table 21 (in Appendix).

4.4 Evaluation of the results

The estimation of the regression model (1.6) used in deriving LRS shows that its statistical
significance is very high. Indeed, its p-value turns out to be 0 and its R-square is 95.2%.
In order to evaluate the results obtained using the LRS model, we shall carry out several
comparisons. We shall compare LRS with the S&P ratings, as well as with the results obtained
using the non-recursive regression model derived from the S&P ratings. We shall further compare
LRS with the scores associated to Moody’s ratings at the end of December 1998, and with those
provided by The Institutional Investor in March 1999. In this analysis, we shall evaluate first the
relative preferences obtained using LAD, and then evaluate the LRS derived from these relative
preferences. The consistency of any two sets of results is measured by their correlation
coefficient.

4.4.1 Evaluation of relative preferences

In order to evaluate the matrix ∆ of relative preferences obtained using LAD, we shall need a
comparable point of reference. A natural benchmark of this sort can be associated to any set of
numerical scores si representing sovereign ratings, by defining the canonical relative preferences
dij to be simply the differences dij = si – s j associated to every pair of countries i and j. In
this paper, we shall compare the LAD relative preferences ∆ij with the canonical relative
preferences dS&Pij , dMij , dIIij , dREGij , and dLRSij obtained respectively from the scores associated
with S&P’s ratings, Moody’s ratings, The Institutional Investor’s scores, the non-recursive
regression model scores, and the logical rating scores. The corresponding matrices of relative
preferences will be denoted dS&P , dM , dII , dREG , and dLRS respectively. The correlation levels
are shown in Table 10, where the symmetric correlation matrix is filled out completely, in order
to make comparisons easier.

Table: Correlation levels between relative preference matrices

dS&P dM dII dREG ∆ DLRS

dS&P 100% 98.01% 96.18% 95.73% 93.21% 95.54%
dM 98.01% 100% 96.31% 95.00% 92.89% 95.20%
dII 96.18% 96.31% 100% 94.16% 91.82% 94.11%
dREG 95.73% 95.00% 94.16% 100% 95.91% 98.29%
∆ 93.21% 92.89% 91.82% 95.91% 100% 97.57%
LRS
d 95.54% 95.20% 94.11% 98.29% 97.57% 100%
RRR 08-2004 PAGE 25

The high levels of correlation show that the relative preferences obtained using LAD are in a
surprisingly good agreement both with the ratings of S&P and those of the other agencies, as well
as with the non-recursive regression scores. A comparison of the logical rating scores with the
LAD relative preferences shows the clear superiority of the former, in view of their higher degree
of agreement with the other canonical relative preferences.

4.4.2 Evaluation of logical rating scores

We shall analyze now the correlation levels between the logical rating scores and the scores
associated with the ratings of S&P, Moody and The Institutional Investor, as well as those
provided by the non-recursive regression model (columns 2, 10 and 11, respectively, in Table 21
in the Appendix). It can be shown (Appendix) that these correlation levels are exactly identical to
those between the corresponding canonical relative preference matrices, which are presented in
Table 10.
Several valuable conclusions can be derived from the correlation levels reported in Table 10.
First, the high levels of correlation between the scores show that both the logical rating scores and
those provided by the non-recursive regression model are very good approximations of the S&P
ratings, as well as of those provided by other rating agencies.
Second, it is remarkable that, in spite of the entirely different nature of the LRS and the non-
recursive regression techniques, their results have an extremely high correlation (98.29%).
Moreover, the correlation between one of these two scores and any individual agency rating, is
almost identical to the correlation between the other score and the rating of that agency.
This similarity is striking due to several essential distinctions between LRS and the non-
recursive regression technique. The first distinction is that the regression technique assumes that
the differences between consecutive ratings are the same, while the LRS is not based on any
assumption about the magnitude of the differences. The second – and perhaps most striking
distinction – concerns the mathematical techniques used for deriving the two scores: the non-
recursive regression score is obtained through a classical statistical technique, while the LRS is
derived through combinatorial-logical concepts and techniques.
The third distinction concerns the mathematical functions describing the relationship between
the independent variables and the scores: while the non-recursive regression scores depend
linearly on the variables, the LRS depends on them in a complex nonlinear fashion, through the
intermediary constructs of logical patterns.
The most important conclusion derived from this discussion is the fact that the almost identical
results, provided by two techniques of essentially different natures, strongly reinforce each other.

4.4.3 Discrepancies between S&P and LRS

It has been seen in the previous section that the LRS and the S&P ratings are in close
agreement. However, since the logical rating scores and the S&P ratings are not expressed on the
same scale, the comparison of the two scores of an individual country presents a challenge. In
order to overcome this difficulty, we shall apply a linear transformation to the LRS, which brings
them to the same scale as the S&P ratings. This is accomplished by determining the coefficients a
and c for the transformation a*βi + c of the LRS βi in such a way that the mean square difference
between the transformed LRS and the S&P ratings is minimized. Obviously, these coefficients
can be determined by simple linear regression. As a result of this transformation, the LRS become

25
PAGE 26 RRR 08-2004

directly comparable with the S&P ratings. Clearly, the consistency of (i.e., the correlation
coefficient between) the LRS and S&P ratings is not affected by this transformation.
Using formula (1.6), we then compute the confidence intervals for the transformed LRS of
each country in the dataset. In 1998, five countries have a S&P rating that does nor fall within the
confidence interval of the transformed LRS. Columbia appears to be too favorably rated by S&P,
while Hong-Kong, Malaysia, Pakistan and Russia appear to be rated too harshly by S&P. As
explained in Sections 3.1.3 and 3.1.4, the evolution of the S&P ratings for Columbia, Pakistan
and Russia is in agreement with the 1998 LRS of these countries, and underlines the prediction
capability of the LRS model. It is remarkable that the evolution of the S&P ratings of Malaysia
and Hong-Kong is also in agreement with their 1998 LRS. Indeed, both Malaysia and Hong-
Kong have been upgraded shortly thereafter, the former moving from BBB- to BBB in November
1999, and the latter from A to A+ in February 2001.

4.5 Extendability of the logical rating scores model

As it has been seen above, the derivation of the LRS model consists of two stages. In the first
stage, LAD is applied to derive patterns of superiority of one country over another and combine
them into a discriminant that provides numerical value of the relative preferences. In the second
stage, the relative preferences are “integrated” into logical rating scores, by using classical linear
regression.
The LRS model makes possible the calculation of logical rating scores on the basis of new
data. Indeed, in order to obtain logical rating scores for the new data all that has to be done is to
recalculate the relative preferences using the existing LAD discriminant, and then to rerun linear
regression in order to “reintegrate” the new relative preferences into the new logical rating scores.
Clearly, the analysis of new data by the LRS model is carried out independently of any possible
new ratings of S&P, since the LAD discriminant underlying the LRS model has already been
constructed and remains valid.
The remarkable feature of the LRS model is that for important cases of data perturbation,
including those resulting from temporal changes in the values of the independent variables, or the
addition of previously non-rated countries, the new logical rating scores maintain to a large extent
their validity. In order to illustrate this idea, we shall show below that the LRS model developed
above on the basis of the 1998 data can be successfully applied on the one hand for deriving LRS
scores based on 1999 data, and on the other hand for deriving LRS scores of previously non-rated
countries.

4.5.1 Temporal changes in data

In this section, we shall construct logical rating scores based on the 1999 data, in order to
evaluate the robustness and consistency of the LRS model. This goal will be accomplished by
comparing the LRS obtained in this way with Standard and Poor’s 1999 ratings. The logical
ratings scores for the 1999 data are given in Appendix (Table 21, columns 8).
It has been seen in Section 4.4.2 that the scores provided by the two models (the non-recursive
regression and the LRS models) built on the 1998 Standard & Poor ratings when applied to the
1998 data are strikingly similar. In this section, we compare the results of the same two methods
applied to the 1999 data.
RRR 08-2004 PAGE 27

Table 11: Correlation levels between relative preference matrices

dS&P dREG ∆ dLRS
dS&P 100% 94.74% 91.61% 94.12%
dREG 94.74% 100% 94.46% 97.50%
∆ 91.61% 94.46% 100% 96.48%
LRS
d 94.12% 97.50% 96.48% 100%

The first conclusion resulting from the high levels of pairwise correlations between the S&P
1999 ratings, the relative preferences given by the LAD discriminant, and the canonical relative
preferences corresponding to non-recursive regression scores and LRS, is that both the LRS and
the non-recursive regression model have a very strong temporal stability. The second conclusion
is that the logical rating scores are superior to the relative preferences given by the LAD
discriminant.
The major conclusion resulting from this table is that the logical rating and the non-recursive
regression scores remain strikingly similar, and each of them provides a strong reinforcement of
the other.

4.5.2 Discrepancies

Using formula (1.6), we compute the confidence intervals for the transformed LRSe of each
country in the dataset. Applying the 1998 LRS model to the 1999 data, we see that only two
countries (Russia and Hong-Kong) have S&P ratings that are outside the confidence intervals of
the corresponding transformed LRS. These two countries appear to be rated too harshly by S&P.
As explained in Sections 3.1.4 and 4.4.3, the evolution of the S&P ratings for Hong-Kong and
Russia is in agreement with the 1999 LRS of these countries.

4.5.3 Previously non-rated countries

The availability of the LAD discriminant, which does not involve in any way the previous
years’ S&P ratings, makes it possible to rate previously non-rated countries in the following way.
After calculating first the attribute values of all the pseudo-observations involving the new
countries to be evaluated, the relative preferences are to be calculated for these pseudo-
observations, and the resulting columns and rows have to be added to the matrix of relative
preferences. The new LRS for all the countries (new and old) should then be determined by
running the multiple linear regression model (1.12).
In order to evaluate the capability of LRS to correctly predict S&P ratings, we compare the
LRS predicted as described above, with the S&P ratings when they first become available. The
direct comparison between these ratings is carried out using the linear transformation described in
Section 4.4.3.
Using formula (1.6), we compute the confidence intervals for the transformed LRS of four
countries never rated by S&P by December 1998. It appears that our predictions for three of them
(Guatemala, Jamaica and Papua New Guinea) correspond perfectly to the first time (subsequent)
S&P ratings. The comparison between the LRS and the first S&P rating (SD) given in July 2000
for the fourth country (Ecuador) shows that S&P rated it much too harshly, since one month later
S&P significantly raised its rating to B-, thus fully justifying the LRS prediction.

27
PAGE 28 RRR 08-2004

4.6 Cross-validation of relative preferences

Since the role of the LAD discriminant is to capture the rules of country creditworthiness,
which are implicit in the S&P ratings, it is the most important component of the LRS
methodology. As any learning procedure, LAD can be susceptible to overfitting, i.e., adapting so
well to the training data to allow random noise to influence the model. If this happens, the
performance of the resulting model can be excellent on the training data, but can perform very
poorly on new observations.
To test whether the LAD discriminant is affected by overfitting, we shall use – similarly to
Section 3.1.2– the commonly used statistical technique of cross-validation. The particular type
of cross-validation technique used here is known as “jackknife” (Quenouille, 1949,1956) or
“leave-one-out”. In broad terms, the jackknife technique consists of removing from the dataset
one observation at a time, learning a model from all the remaining observations, evaluating the
resulting model on the removed observation; and then repeating these steps for each observation
in the dataset. If the predicted evaluations are “close to” the actual values of the observations, as
well as in-the-sample evaluations, then the model is not affected by overfitting.
Each step of the implementation of jackknife used here eliminates from the dataset all the
pseudo-observations involving a certain country, derives the LAD discriminant on the basis of the
remaining pseudo-observations, and then uses this discriminant to evaluate the relative preference
for every removed pseudo-observation. This provides a column of relative preferences of pseudo-
observations involving the country singled out at this step. After repeating this procedure for
every country in the dataset, we combine the obtained columns into a matrix of relative
preferences denoted by ∆ JK. Then we construct the logical rating scores on the basis of ∆ JK and
denote the corresponding matrix of canonical relative preferences by dJKLRS.
Table12: Correlation between cross-validated relative preferences
dS&P ∆ dLRS ∆ JK dJKLRS
S&P
d 100% 93.21% 95.54% 92.98% 95.26%
∆ 93.21% 100% 97.57% 96.48% 96.36%
dLRS 95.54% 97.57% 100% 96.35% 96.89%
∆ JK 92.98% 96.48% 96.35% 100% 97.43%
dJKLRS 95.26% 96.36% 96.89% 97.43% 100%

The correlation levels presented in Table 12 indicate that the matrices of relative preferences
∆JK and dJKLRS obtained through the jackknife procedure are highly correlated with (i) the
canonical relative preferences corresponding to the S&P ratings dS&P , and (ii) the in-the-sample
relative preferences ∆, as well as the corresponding logical rating scores dLRS . This shows that ∆
and dLRS are not affected by overfitting.

5 Concluding Remarks
The central objective of this paper was to develop transparent, consistent, self-contained, and
stable country risk rating systems, closely approximating the country risk ratings provided by a
major rating agency (Standard & Poor). We proposed in this paper two models that achieve the
stated objectives, the first one utilizing the classical econometric technique of multiple linear
RRR 08-2004 PAGE 29

regression, and the second one using the combinatorial-logical technique of Logical Analysis of
Data.
The proposed models are highly accurate, having a 95.5% correlation level with the actual
S&P ratings and almost equally high correlations with Moody’s and The Institutional Investor.
The models avoid overfitting, as demonstrated by the 99.1% correlation between in- and out-of-
sample rating predictions, calculated by k-fold cross-validation. The proposed models are
transparent since they make the role of the economic-financial and political variables explicit.
The proposed models are distinguished from the rating models in the existing literature by
their self-contained nature, i.e., by their non-reliance on any information derived from lagged
ratings. Therefore, the high level of correlation between predicted and actual ratings cannot be
attributed to the reliance on lagged ratings, and is a reflection of the relevance and predictive
power of the independent variables included in these models. The significant advantage of the
non-recursive nature of the proposed models is their applicability to not-yet-rated countries.
The consistency of the proposed models is illustrated by the fact that the few discrepancies
between the S&P ratings and those of the proposed models were resolved by subsequent changes
in S&P’s ratings. The stability of the constructed non-recursive models is shown in two ways: by
predicting 1999 ratings using the non-recursive models derived from the 1998 dataset applied to
the 1999 data, and – most importantly -- by successfully predicting the ratings of several
previously non-rated countries, i.e., in full agreement with subsequent ratings of those countries
by S&P.
The confidence in the results and in the validity of both models is strongly reinforced by the
fact that the traditional linear regression model and the qualitatively different combinatorial-
logical model produce almost identical results.
The significance of the results of this paper is further confirmed in a forthcoming study
(Hammer et al., 2004), in which the LAD discriminant constructed in this paper is used for
deriving a partial order representing the creditworthiness of countries, and then extending this
partial order to a new rating system. The comparison of the results of the present and the
forthcoming studies shows that – in spite of yet another qualitatively different data analysis
approach -- the predicted ratings are strongly correlated, at the surprising level of 98-99%.

29
PAGE 30 RRR 08-2004

References

Altman E.I. and Saunders A. 1997. Credit Risk Measurement: Developments over the Last 20
Years. Journal of Banking and Finance 21 (11-12), 1721-1742.
Aylward L. and Thorne R. 1998. Countries’ Repayments Performance Vis-à-Vis the IMF”, IMF Staff
Papers 45 (4), 595-619.
Bhatia A.V. 2002. Sovereign Credit Risk Ratings: An Evaluation. IMF Working Pape WP/
03/170.
Boros E., Hammer P.L., Ibaraki T. and Kogan A. 1997. Logical Analysis of Numerical Data,
Mathematical Programming 79, 163-190.
Bouchet M.H., Clark E. and Groslambert B. 2003. Country Risk Assessment: A Guide to
Global Investment Strategy. John Wiley & Sons Ltd. Chichester, England.
Bourke P. and Shanmugam B. 1990. An Introduction to Bank Lending. Addison-Wesley
Business Series.
Breiman L. and Spector P. 1992. Submodel Selection and Evaluation in Regression: The X-Random
Case. International Statistical Review 60, 291-319.
Brewer T.L. and Rivoli P. 1990. Politics and Perceived Country Creditworthiness in
International Banking. Journal of Money, Credit and Banking 22, 357-369.
Brewer T.L. and Rivoli P. 1997. Political Instability and Country Risk. Global Finance
Journal 8 (2), 309-321.
Cantor R. and Packer F. 1996. Determinants and Impact of Sovereign Credit Ratings. FRBNY
Economic Policy Review, 37-53.
Citron J.T. and Neckelburg G. 1987. Country Risk and Political Instability. Journal of
Development Economics 25, 385-395.
Claessens S. and Embrechts G. 2002. Basel II, Sovereign Ratings and Transfer Risk: External
versus Internal Ratings. Presentation at the Basel II: An Economic Assessment, Bank for
International Settlements, Basel, Switzerland.
Cook W.D. and Hebner K.J. 1993. A Multicriteria Approach to Country Risk Evaluation:
With an Example Employing Japanese Data. International Review of Economics and Finance 2
(4), 327-348.
Coplin W.D. and O’Leary M.K. 19983. Political Risk Yearbook, The PRS Group, New York.
Cosset J.C. and Roy J. 1991. The Determinants of Country Risk Ratings. Journal of
International Business Studies 22, 135-142.
Cosset J.C., Siskos Y. and Zopounidis C. 1992. Evaluating Country Risk: A Decision Support
Approach. Global Finance Journal 3, 79-95.
Crama Y., Hammer P.L. and Ibaraki T. 1988. Cause-Effect Relationships and Partially
Defined Boolean Functions. Annals of Operations Research 16, 299-326.
Cruces J.J. 2003. Statistical Properties of Sovereign Credit ratings. Working Paper. University
of San Andres, Buenos Aires.
Doumpos M. and Zopounidis C. 2001. Assessing Financial Risks Using a Multicriteria
Sorting Procedure: The Case of Country Risk Assessment. Omega 29, 97-109.
RRR 08-2004 PAGE 31

Easton S.T. and Rockerbie D.W. 1999. What’s in a Default ? Lending to LDCs in the Face of
Default Risk. Journal of Development Economics 58 (2), 319-332.
Eaton J., Gersovitz M. and Stiglitz J.E. 1986, The Pure Theory of Country Risk. European
Economic Review 30, 481-513.
Efron B. 1982. The Jackknife, the Bootstrap and Other Resampling Plans. Philadelphia, SIAM.
Erb C.B., Harvey C.R. and Viskanta T.E. 1996a. Expected Returns and Volatility in 135
Countries. Journal of Portfolio Management, 46-58.
Erb C.B., Harvey C.R. and Viskanta T.E. 1996b. Political Risk, Economic Risk and Financial
Risk. Financial Analysts Journal, 29-46.
Estrella A. 2000. Credit Ratings and Complement Sources of Credit Quality Information. Basel
Committee on Banking Supervision Working Paper 3.
Feder G. and Uy V.U. 1985. The Determinants of International Creditworthiness and their
Policy Implications. Policy Modeling 7 (1), 133-156.
Ferri G., Liu L-G. and Stiglitz J. 1999. The Procyclical Role of Rating Agencies: Evidence
from the East Asian Crisis. Economic Notes 3, 335-355.
Hammer P.L. 1986. Partially Defined Boolean Functions and Cause-Effect Relationships.
International Conference on Multi-Attribute Decision Making Via OR-Based Expert Sytems.
University of Passau, Passau, Germany.
Hammer P.L., Kogan A. and Lejeune M.A. 2004. Modeling Country Risk Ratings Using
Partial Orders. RUTCOR Technical Report 1-2004, Rutgers University, New Brunswick, New
Jersey.
Haque N.U., Kumar M.S., Mark N. and Mathieson D. 1996. The Economic Content of
Indicators of Developing Country Creditworthiness. International Monetary Fund Working
Paper 43 (4), 688-724.
Haque N.U., Kumar M.S., Mark N. and Mathieson D. 1998. The Relative Importance of
Political and Economic Variables in Creditworthiness Ratings. International Monetary Fund
Working Paper 46, 1-13.
Haque N.U., Mark N. and Mathieson D. 1997. Rating the Raters of Country Creditworthiness.
Finance & Development 34, 10-13.
Hjorth J.S.U. 1994. Computer Intensive Statistical Methods Validation, Model Selection, and
Bootstrap. London, Chapman & Hall.
Hu Y.-T., Kiesel R. and Perraudin W. 2002. The Estimation of Transition Matrices for
Sovereign Credit Ratings. Journal of Banking and Finance 26 (7), 1383-1406.
Hurvich C.M. and Tsai C.-L. 1989. Regression and Time Series Model Selection in Small Samples.
Biometrika 76, 297-307.
Jüttner J.D. and McCarthy J. 2000. Modeling a Rating Crisis. Macquarie University,
Unpublished, Sidney, Australia, http://www.econ.mq.edu.au/staff/djjuttner/SOVEIG2.pdf.
Kaminsky G. and Schmukler S.L. 2002. Emerging Market Instability: Do Sovereign Ratings Affect
Country Risk and Stock Returns? World Bank Economic Review 16, 171-195.
Kaufmann D., Kraay A. and Zoido-Lobaton P. 1999a. Aggregating Governance Indicators.
World Bank Policy Research Department Working Paper 2195,
http://www.worldbank.org/wbi/governance/pdf/agg_ind.pdf.
Kaufmann D., Kraay A. and Zoido-Lobaton P. 1999b. Governance Matters. World Bank

31
PAGE 32 RRR 08-2004

Policy Research Department Working Paper 2196,

http://www.worldbank.org/wbi/governance/pdf/govmatrs.pdf.
Kaufmann D., Kraay A. and Zoido-Lobaton P. 2002. Governance Matters II: Updated
Indicators for 2000/01. World Bank Policy Research Department Working Paper 2772,
http://ideas.repec.org/p/wop/wobago/2772.html.
Kräussl R. 2000. Sovereign Ratings and Their Impact on Recent Financial Crises. Working
paper, Center For Financial Studies. Frankfurt.
Kunczik M. 2000. Globalization: News Media, Images of Nations and the Flow of International Capital
with Special Reference to the Role of Rating Agencies, Paper presented at the IAMCR Conference,
Singapore, 1-49.

Larrain G., Reisen H. and von Maltzan J. 1997. Emerging Market Risk and Sovereign Credit
Ratings. OECD Development Center, 1-30.
Lee S.H. 1993. Relative Importance of Political Instability and Economic Variables on
Perceived Country Creditworthiness. Journal of International Business Studies 24, 801-812.
Levich R., Reinhart C., and Majnoni G. 2002. Ratings, Rating Agencies and the Global
financial System. Kluwer Academic Press. New York.
Manasse P., Roubini N. and Schimmelpfennig A. 2003. Predicting Sovereign Debt Crises.
IMF Working Pape WP/ 03/221.
Mathieson D. and Schinasi G. 1999. International Capital Markets: Developments, Prospects,
and Key Policy Issues. IMF Series: World Economic and Financial Surveys.
Mauro P. 1993. Essays on Country Risk, Asset Markets and Economic Growth. Harvard
University Dissertation.
Mauro P. 1998. Corruption and the Composition of Government Expenditure. Journal of
Public Economics 69 (2), 263-279.
Monfort B. and Mulder C. 2000. Using Credit Ratings for Capital Requirements on Lending
to Emerging Market Economies: Possible Impact of a New Basel Accord. IMF Working Paper
WP/00/69.
Moody’s. 2001. Moody’s Country Credit Statistical Handbook, First Edition, New York..
Mulder C. and Perelli R. 2001. Foreign Currency Credit Ratings for Emerging Market
Economies. IMF Working Paper WP/01/191.
Mulekar M.S. and Mishra S.N. 2000, Confidence Interval Estimation of Overlap: Equal
Means Case, Computational Statistics & Data Analysis 34, 121-137.
Bilson C. M., Brailsford T.J. and Hooper V.J. 2001. Selecting Macroeconomic Variables as
Explanatory Factors of Emerging Stock Market Returns. Pacific-Basin Finance Journal 9 (4),
401-426 .
Neter J., Kutner M.H., Nachtsheim C.J. and Wasserman W. 1996. Applied Linear Statistical
Models. Fourth Edition. Irwin/McGraw-Hill.
Nickell P., Perraudin W. and Varotto S. 2000. Stability of Rating Transitions. Journal of
Banking and Finance 24, 203-227.
Nordal K.B. 2001. Country Risk, Country Risk Indices and Valuation of FDI: A Real Options
Approach. Emerging Markets Review 2, 197-217.
RRR 08-2004 PAGE 33

Oral M., Kettani O., Cosset J.-C. and Daouas M. 1992. An Estimation Model for Country
Risk Rating. International Journal of Forecasting 8 (4), 583-593.
Quenouille, M. 1949. Approximate Tests of Correlation in Time Series. Journal of the Royal
Statistical Society, Series B 11, 18-84.
Quenouille, M. 1956. Notes on Bias in Estimation. Biometrika 43, 353-360.
Reinhart C.M. 2002. Default, Currency Crises, and Sovereign Credit Ratings. World Bank
Economic Review 16, 151-170.
Reisen H. and von Maltzan J. 1998. Sovereign Credit Ratings, Emerging Market Risk and
Financial Market Volatility. Intereconomics 33 (2), 73-82.
Reisen H. and Von Maltzan J. 1999. Boom and Bust and Sovereign Ratings. International
Finance 2 (2), 273-293.
Saini K.G. and Bates P.S. 1984. A Survey of the Quantitative Approaches to Country Risk Analysis.
Journal of Banking and Finance 8, 341-356.
Shao J. 1993. Linear Model Selection by Cross-Validation. Journal of the American
Statistical Association 88, 486-494.
Shao J. and Tu D. 1995. The Jackknife and Bootstrap. New York, Springer-Verlag.
Simpson J. 2000. The World Trade Organisation and Basel Accord Membership and
Compliance in Relation to International Banking Risk. Curtin University Working Paper 35.
Somerville R.A. and Taffler R.J. 1995. Banker Judgment versus Formal Forecasting Models :
The Case of Country Risk Assessment. Journal of Banking and Finance 19 (2), 281-297.
Standard & Poor. 1999. Sovereign Ratings Display Stability over Two Decades. Credit Week.
Standard & Poor. 2000. Ratings Performance. Standard & Poor. New York.
Standard & Poor. 2001. Sovereign Ratings History since 1975. New York.
Sy A.N.R. 2003. Rating the Rating Agencies: Anticipating Currency Crises or Debt Crises?
IMF Working Paper WP/ 03/122.
Tang J.C.S. and Espinal C.G. 1989. A Model to Assess Country Risk. OMEGA: International
Journal of Management Sciences 17 (4), 363-367.
World Bank. 2000a. Global Development Finance 2000. Washington D.C.
World Bank. 2000b. World Development Indicators 2000. Washington D.C.

33
PAGE 34 RRR 08-2004

Appendix
Table 13:Economic variables and literature

Variable Literature
Consumer price index Larrain et al. (1997), Hu et al. (2002)
Credit claims on central government
Monfort and Mulder (2000)
growth rate
Haque et al. (1996,1998), Larrain et al. (1997), Brewer and Rivoli
Current account balance / GDP (1990), Doumpos et al. (2001), Cosset et al.(1992), Cook and
Hebner (1993), Tang and Espinal (1990)
Larrain et al. (1997), Feder and Uy (1985), Aylward and Thorne
(1998), Dailami and Leipziger (1997), Cosset and Roy (1991),
Debt12 / exports Cosset et al. (1992), Cantor and Packer (1996), Lee (1993),
Doumpos and Zopounidis (2001), Monfort and Mulder (2000), Hu
et al. (2002), Mulder and Perelli (2002)
Haque et al. (1996, 1998), Feder and Uy (1985), Lee (1993),
Debt13 / GDP Brewer and Rivoli (1990), Aylward and Thorne (1998), Doumpos
and Zopounidis (2001), Cook and Hebner (1993), Hu et al. (2002)
Debt / reserves Monfort and Mulder (2000), Manasse et al. (2003)
Dependence on oil exportation Feder and Uy (1985)
Domestic investment / GDP Larrain et al. (1997), Monfort and Mulder (2000)
Exports / GDP Aylward and Thorne (1998)
Exports concentration Feder and Uy (1985)
Haque et al. (1996,1998), Feder and Uy (1985), Doumpos and
Exports growth rate Zopounidis (2001), Cosset et al.(1992), Monfort and Mulder
(2000)
Exports variability Cosset et al.(1992)
Exports vulnerability to external
Feder and Uy (1985)
shocks
Larrain et al. (1997}, Brewer and Rivoli (1990), Monfort and
External debt / GDP
Mulder (2000), Manasse et al. (2003)
Larrain et al. (1997), Cantor and Packer (1996), Lee (1993),
Fiscal balance14
Monfort and Mulder (2000) , Cook and Hebner (1993)
Foreign investment policy Cook and Hebner (1993)
Haque et al. (1996,1998), Larrain et al. (1997), Feder and Uy
(1985), Cantor and Packer (1996), Doumpos and Zopounidis
GDP15 growth rate
(2001), Monfort and Mulder (2000), Cook and Hebner (1993), Hu
et al.(2002)
Dailami and Leipziger (1997), Erb et al. (1997), Feder and Uy
GDP per capita (1985), Cosset et al. (1991, 1992), Larrain et al. (1997), Monfort
and Mulder (2000), Tang and Espinal (1990)
GDP per capita growth rate Lee (1993), Haque et al. (1996), Aylward and Thorne (1998)
Easton and Rockerbie (1999), Cosset et al. (1991, 1992),
Gross investment / GDP
Doumpos and Zopounidis (2001)

12
Same as for footnote 6.
13
The word “debt” can encompass foreign, total, debt service or external debt, depending on authors.
14
Central government spending / GDP, domestic public debt / GDP and are used as a proxy for this variable.
15
Note that authors use GDP as well as GNP.
RRR 08-2004 PAGE 35

Imports / GDP Haque et al.(1996), Aylward and Thorne (1998)

Imports growth rate Doumpos and Zopounidis (2001)
Income velocity of money
Doumpos and Zopounidis (2001)
(GDP/M2)
Indicator for economic development Cantor and Packer (1996)
Haque et al. (1996,1998), Larrain et al. (1997), Dailami and
Leipziger (1997), Erb et al. (1997), Aylward and Thorne (1998),
Inflation rate
Cantor and Packer (1996), Doumpos and Zopounidis (2001),
Monfort and Mulder (2000)
Haque et al. (1996,1998), Easton and Rockerbie et al. (1999),
Dailami and Leipziger (1997), Feder and Uy (1985), Aylward and
International reserves / imports Thorne (1998), Lee (1993), Doumpos and Zopounidis (2001),
Cosset et al.(1992), Monfort and Mulder (2000), Tang and
Espinal (1990), Hu et al.(2002)
Long-term debt / GDP Easton and Rockerbie et al. (1999)
Oil exportation16 Feder and Uy (1985)
Haque et al. (1996,1998), Larrain et al.(1997), Monfort and
Real exchange rate
Mulder (2000), Cook and Hebner (1993)
Savings / GDP Larrain et al. (1997)
Short-term debt / reserves Dailami and Leipziger (1997)
Short-term debt / total debt Monfort and Mulder (2000)
Haque et al. (1996,1998), Easton and Rockerbie (1999), Feder and
Terms of trade Uy (1985), Doumpos and Zopounidis (2001), Monfort and Mulder
(2000)
Trade openness Easton and Rockerbie (1999)
Treasury bill rate Haque et al. (1998), Monfort and Mulder (2000)

Table 14: Political variables and literature

Variable Literature
Anti-governmental demonstrations Haque et al. (1998)
Armed conflicts (or riots) Haque et al. (1998), Brewer and Rivoli (1990), Cook and Hebner
(1993)
Assassination Haque et al. (1998)
Corruption Mauro (1993)
Coups Haque et al. (1998)
General strikes Haque et al. (1998)
Guerilla warfare Haque et al. (1998)
Influence of the middle class Mauro (1993), Cook and Hebner (1993)
Legal system Mauro (1993)
Major government crises Haque et al. (1998)
Political change Mauro (1993), Brewer and Rivoli (1990)
Political legitimacy Brewer and Rivoli (1990)
Political stability Brewer and Rivoli (1990, 1997), Feder and Uy (1985), Citron and
Neckelburg (1987), Mauro (1993), Lee (1993), Cosset et al.(1992),

16
Represented by a dummy variable.

35
PAGE 36 RRR 08-2004

Cook and Hebner (1993), Manasse et al. (2003)

Probability of opposition group Mauro (1993)
takeover
Purges Haque et al. (1998)
Red tape, bureaucracy Mauro (1993)
Relationships with neighboring Mauro (1993)
countries
Revolutions Haque et al. (1998)
Social Stability Cook and Hebner (1993)
Stability of labor Mauro (1993)
Terrorism Mauro (1993)

Table 15: Standard & Poor’s country rating system

Level Description
An obligor rated AAA has extremely strong capacity to meet its financial
AAA
commitments. AA is the highest issuer credit rating assigned by S&P.
An obligor rated AA has very strong capacity to meet its financial
INVESTMENT

AA
commitments. It differs from the highest rated obligors only in small degree.
RATING

An obligor rated A has strong capacity to meet its financial commitments but is
A somewhat more susceptible to the adverse effects of changes in circumstances
and economic conditions than obligors in higher-rated categories.
An obligor rated BBB has adequate capacity to meet its financial commitments.
However, adverse economic conditions or changing circumstances are more
BBB
likely to lead to a weakened capacity of the obligor to meet its financial
commitments.
An obligor rated BB is less vulnerable in the near term than other lower-rated
SPECULATIVE

obligors. However, it faces major ongoing uncertainties and exposure to

BB
adverse business, financial, or economic conditions which could lead to its
RATING

inadequate capacity to meet financial commitments.

An obligor rated B is more vulnerable than the obligors rated BB, but, at the
time of the rating, it has the capacity to meet financial commitments. Adverse
B
business, financial, or economic conditions could likely impair its capacity or
willingness to meet financial commitments.
An obligor rated CCC is vulnerable at the time of the rating, and is dependent
CCC upon favorable business, financial, and economic conditions to meet financial
DEFAULT

commitments.
RATING

CC An obligor rated CC is highly vulnerable at the time of the rating.

An obligor rated C is vulnerable to nonpayment at the time of the rating and is
C dependent upon favorable business, financial, and economic conditions to meet
financial commitments.
D An obligator rated D is predicted to default.
SD An obligator rated SD (selected default) is presumed to be unwilling to repay.
RRR 08-2004 PAGE 37

Table 16: Standard & Poor’s country ratings (end of December, 1998)
We have converted the Standard & Poor rating scale (columns 1 and 4) into a numerical scale
(columns 2 and 5). Such a conversion is not specific to us. Bouchet et al. (2003), Estrella (2000),
Ferri et al.(2001), Kräussl (2000), Monfort and Mulder (2000), Mulder and Perelli (2001), Sy
[2003] proceed similarly. Moreover, Bloomberg, a major provider of financial data services,
developed a standard cardinal scale for comparing Moody’s, S&P and Fitch-BCA ratings
(Kaminsky and Schmukler, 2002). A higher numerical value denotes a higher probability of
default. The numerical scale is referred to in this paper as Standard & Poor’s preorder.

Rating Preorder Country Rating Preorder Country

AAA 21 AUSTRIA BBB- 12 CROATIA
AAA 21 FRANCE BBB- 12 EGYPT
AAA 21 GERMANY BBB- 12 LITHUANIA
AAA 21 JAPAN BBB- 12 MALAYSIA
AAA 21 NETHERLANDS BBB- 12 POLAND
AAA 21 NORWAY BBB- 12 THAILAND
AAA 21 SINGAPORE BBB- 12 TUNISIA
AAA 21 SWITZERLAND BBB- 12 URUGUAY
AAA 21 UNITED KINGDOM BB+ 11 KOREA
AAA 21 UNITED STATES BB+ 11 PANAMA
AA+ 20 BELGIUM BB+ 11 PHILIPPINES
AA+ 20 CANADA BB+ 11 SLOVAK REPUBLIC
AA+ 20 DENMARK BB+ 11 SOUTH AFRICA
AA+ 20 IRELAND BB+ 11 TRINIDAD AND TOBAGO
AA+ 20 NEW ZEALAND BB 10 ARGENTINA
AA+ 20 SWEDEN BB 10 COSTA RICA
AA 19 AUSTRALIA BB 10 EL SALVADOR
AA 19 FINLAND BB 10 INDIA
AA 19 ITALY BB 10 MEXICO
AA 19 PORTUGAL BB 10 MOROCCO
AA 19 SPAIN BB 10 PERU
A+ 17 CYPRUS BB- 9 BOLIVIA
A+ 17 ICELAND BB- 9 BRAZIL
A+ 17 MALTA BB- 9 JORDAN
A 16 HONG KONG BB- 9 LEBANON
A 16 SLOVENIA BB- 9 PARAGUAY
A- 15 CHILE B+ 8 DOMINICAN REPUBLIC
A- 15 CZECH REPUBLIC B+ 8 KAZAKHSTAN
A- 15 ISRAEL B+ 8 VENEZUELA
BBB+ 14 CHINA B 7 TURKEY
BBB+ 14 ESTONIA B- 6 ROMANIA
BBB 13 GREECE CCC+ 5 INDONESIA
BBB 13 HUNGARY CCC- 3 RUSSIA
BBB 13 LATVIA CC 2 PAKISTAN
BBB- 12 COLUMBIA

37
PAGE 38 RRR 08-2004

Table 17: Ratings with the non-recursive regression model approach

Non-recursive regression model: predicted
country risk ratings and 10-fold cross-
validation Country risk ratings Predicting 1999 Predicting 1999
based on economic country risk ratings country risk ratings
Predicted country Cross-validation
variables using 1998 model using 1999 model
Countries risk rating country risk rating
(in-the-sample) (out-of-the-sample)
Argentina 12.09 14.01 13.25 14.19 13.32
Australia 19.92 20.85 20.23 21.37 18.19
Austria 20.62 19.51 20.13 20.52 18.43
Belgium 17.91 16.55 18.32 18.44 17.53
Bolivia 8.88 8.84 8.73 8.88 7.95
Brazil 9.62 10.17 10.37 9.77 7.96
Canada 20.20 20.18 20.53 20.85 18.90
Chile 14.99 14.63 14.80 15.15 12.96
China 13.37 13.06 11.44 13.37 12.47
Colombia 17.61 18.88 18.45 18.59 17.63
Costa Rica 8.43 7.38 8.84 8.74 8.11
Croatia 11.82 12.06 12.36 12.42 9.28
Cyprus 11.89 11.62 11.52 11.36 10.49
Czech Republic 17.16 17.11 17.46 18.14 16.68
Denmark 15.04 15.03 14.45 15.44 13.44
Dominican Rep 20.30 20.30 20.13 20.73 18.45
Egypt 8.82 8.78 8.69 9.65 8.47
El Salvador 10.73 10.38 10.01 10.71 10.37
Estonia 9.99 9.88 9.60 10.48 9.72
Finland 12.07 11.92 13.34 12.60 10.82
France 19.28 19.20 19.74 20.00 17.03
Germany 18.42 18.27 18.97 19.18 17.14
Greece 20.34 20.22 20.44 20.96 18.41
Hong-Kong 13.44 13.59 14.25 14.45 12.32
Hungary 13.28 13.01 13.29 13.48 9.33
Iceland 20.39 20.65 21.03 21.40 19.18
India 8.05 7.48 9.40 8.62 6.29
Indonesia 5.28 5.48 6.63 6.75 4.23
Ireland 18.91 19.11 20.09 20.49 18.35
Israel 14.00 14.19 14.53 14.20 14.60
RRR 08-2004 PAGE 39

Italy 17.30 17.52 16.90 17.36 15.78

Japan 20.07 19.85 19.62 20.08 17.58
Jordan 9.95 9.73 11.21 10.57 7.73
Kazakhstan 8.57 8.41 7.92 9.00 6.83
Korea. Rep. 14.36 14.14 14.54 15.39 12.27
Latvia 11.20 10.82 11.53 11.32 9.55
Lebanon 9.86 10.41 10.07 9.34 9.32
Lithuania 10.57 10.03 12.12 10.62 9.03
Malaysia 14.48 15.00 13.83 14.23 11.45
Malta 15.87 15.73 16.03 16.13 13.54
Mexico 10.35 10.15 9.97 10.95 8.40
Morocco 10.38 10.70 10.33 10.20 8.97
Netherlands 20.62 20.56 21.04 21.01 17.85
New Zealand 19.81 19.49 18.96 19.66 16.93
Norway 22.60 22.92 20.72 22.40 21.10
Pakistan 5.67 5.94 7.16 6.24 4.07
Panama 8.87 10.69 10.39 11.52 9.60
Paraguay 7.46 7.55 6.57 7.48 8.45
Peru 10.57 10.88 9.60 10.25 9.81
Philippines 10.58 10.94 9.75 9.94 6.36
Poland 12.95 12.84 12.68 12.90 9.93
Portugal 17.57 17.37 16.87 17.44 14.91
Romania 7.93 8.97 5.49 8.32 5.72
Russia 6.37 7.27 1.59 6.49 4.04
Singapore 20.12 19.58 19.23 18.54 17.08
Slovak Republic 11.03 11.30 11.85 12.74 10.52
Slovenia 14.19 13.80 14.80 15.28 12.87
South Africa 11.35 11.57 12.04 12.87 11.08
Spain 17.66 17.75 18.31 18.24 15.40
Sweden 19.26 19.15 19.42 19.54 17.52
Switzerland 23.69 24.07 22.81 23.76 21.46
Thailand 12.15 12.54 11.94 13.13 10.34
Trinidad & Tob 11.04 10.75 12.61 12.68 10.41
Tunisia 11.59 11.65 13.58 12.92 8.60
Turkey 5.07 2.94 4.81 6.41 4.52
UK 20.27 20.15 20.44 20.77 18.56
United States 21.08 23.05 22.07 23.98 22.90
Uruguay 13.31 13.25 13.63 14.12 10.43
Venezuela 7.44 6.79 6.55 8.42 6.07

39
PAGE 40 RRR 08-2004

Table 18: Moody’s rating system

Levels Meaning Levels Meaning

Highest quality Ba1 Likely to fulfill

SPECULATIVE RATING
Aaa
obligations
Aa1 High quality Ba2
Aa2 Ba3 Ongoing uncertainty
Aa3 B1 High risk obligations
INVESTMENT

A1 Strong payment capacity B2

RATING

A2 B3
Caa Current vulnerability to
A3
default or in default

DEFAULT
RATING
Adequate payment Ca In bankruptcy or default.
Baa1
capacity
Baa2 D
Baa3

Table 19: Standard & Poor’s country risk ratings: average one-year transition rates (1975-1999)

AAA AA A BBB BB B CCC SD

AAA97.45 2.55 0.00 0.00 0.00 0.00 0.00 0.00
AA 0.71 97.14 0.71 0.00 0.71 0.71 0.00 0.00
A 0.00 4.05 93.24 2.70 0.00 0.00 0.00 0.00
BBB 0.00 0.00 5.33 88.00 5.33 1.33 0.00 0.00
BB 0.00 0.00 0.00 7.06 83.53 7.06 0.00 2.35
B 0.00 0.00 0.00 0.00 14.81 81.48 0.00 3.70
CCC 0.00 0.00 0.00 0.00 0.00 33.33 33.33 33.33
Source: Standard & Poor (2000)
RRR 08-2004 PAGE 41

Table 20: Ratings

Non-recursive Non-recursive The Institutional
S&P ratings S&P preorder LRS ratings S&P ratings S&P preorder LRS ratings Moody’s
Countries regression regression Investor ratings
(1998) (1998) (1998) (1999) (1999) (1999) ratings (1998)
ratings (1998) ratings (1999) (1998)
Argentina BB 10 -0.2768 13.913 BB 10 -0.263 14.186 9 42.7
Australia AA 19 -0.0289 20.117 AA+ 20 -0.0128 21.369 19 74.3
Austria AAA 21 -0.0094 19.655 AAA 21 0.0038 20.524 21 88.7
Belgium AA+ 20 -0.0476 17.282 AA+ 20 -0.0439 18.443 20 83.5
Bolivia BB- 9 -0.366 9.132 BB- 9 -0.3518 8.877 8 28
Brazil BB- 9 -0.3744 9.164 B+ 8 -0.4016 9.766 7 37.4
Canada AA+ 20 -0.0241 20.269 AA+ 20 -0.0112 20.850 20 83
Chile A- 15 -0.191 15.285 A- 15 -0.1841 15.155 14 61.8
China BBB+ 14 -0.2159 13.589 BBB 13 -0.224 13.371 15 57.2
Colombia BBB- 12 -0.3854 8.337 BB+ 11 -0.3964 8.738 12 44.5
Costa Rica BB 10 -0.2748 11.586 BB 10 -0.257 12.418 11 38.4
Croatia BBB- 12 -0.297 11.866 BBB- 12 -0.3202 11.360 12 39.03
Cyprus A+ 17 -0.1081 17.711 A 16 -0.1021 18.143 16 57.3
Czech Republic A- 15 -0.2088 15.063 A- 15 -0.1904 15.436 14 59.7
Denmark AA+ 20 -0.048 20.373 AA+ 20 -0.0492 20.729 20 84.7
Dominican Rep B+ 8 -0.3568 8.768 B+ 8 -0.3431 9.650 10 28.1
Egypt BBB- 12 -0.2915 11.247 BBB- 12 -0.3067 10.710 11 44.4
El Salvador BB 10 -0.3379 10.460 BB+ 11 -0.3301 10.482 12 31.2
Estonia BBB+ 14 -0.2518 12.486 BBB+ 14 -0.245 12.602 14 42.8
Finland AA 19 -0.064 19.362 AA+ 20 -0.0458 20.005 21 82.2
France AAA 21 -0.0828 18.262 AAA 21 -0.0614 19.179 21 90.8
Germany AAA 21 -0.001 20.149 AAA 21 0.0126 20.959 21 92.5
Greece BBB 13 -0.2255 13.239 A- 15 -0.1917 14.452 14 56.1
Hong-Kong A 16 -0.017 18.360 A 16 0.0213 18.594 15 61.8
Hungary BBB 13 -0.2442 12.623 BBB 13 -0.247 13.483 13 55.9
Iceland A+ 17 -0.047 20.544 A+ 17 -0.0378 21.400 18 67
India BB 10 -0.4063 8.548 BB 10 -0.3994 8.622 10 44.5
Indonesia CCC+ 5 -0.4576 4.821 CCC+ 5 -0.4316 6.747 6 27.9
Ireland AA+ 20 -0.0179 18.929 AA+ 20 -0.0249 20.491 21 81.8
Israel A- 15 -0.2215 13.934 A- 15 -0.2189 14.200 15 54.3
Italy AA 19 -0.1064 17.066 AA 19 -0.1122 17.362 18 79.1
Japan AAA 21 -0.0604 19.106 AAA 21 -0.0506 20.079 20 86.5
Jordan BB- 9 -0.323 10.532 BB- 9 -0.2818 10.574 9 37.3
Kazakhstan B+ 8 -0.4095 7.715 B+ 8 -0.4048 9.005 9 27.9
Korea. Rep. BB+ 11 -0.2649 12.822 BBB 13 -0.2182 15.386 11 52.7

41
PAGE 42 RRR 08-2004

Latvia BBB 13 -0.3026 11.281 BBB 13 -0.3039 11.316 13 38

Lebanon BB- 9 -0.3223 10.625 BB- 9 -0.3121 9.342 8 31.9
Lithuania BBB- 12 -0.3247 11.255 BBB- 12 -0.3233 10.621 11 36.1
Malaysia BBB- 12 -0.1676 13.589 BBB 13 -0.1712 14.235 12 51
Malta A+ 17 -0.0999 16.302 A 16 -0.2402 16.131 15 61.7
Mexico BB 10 -0.3608 9.548 BB 10 -0.3284 10.951 10 46
Morocco BB 10 -0.2952 10.587 BB 10 -0.2881 10.198 11 43.2
Netherlands AAA 21 0.0251 20.525 AAA 21 0.0337 21.009 21 91.7
New Zealand AA+ 20 0.0001 19.613 AA+ 20 0.0001 19.661 19 73.1
Norway AAA 21 0.0125 22.234 AAA 21 -0.0076 22.399 21 86.8
Pakistan CC 2 -0.4563 5.184 B- 6 -0.4501 6.236 5 20.4
Panama BB+ 11 -0.2712 11.039 BB+ 11 -0.2487 11.522 11 39.9
Paraguay BB- 9 -0.3865 7.920 B 7 -0.4066 7.481 7 31.3
Peru BB 10 -0.3536 10.386 BB 10 -0.3644 10.250 9 35
Philippines BB+ 11 -0.3242 9.595 BB+ 11 -0.349 9.940 11 41.3
Poland BBB- 12 -0.2772 12.718 BBB 13 -0.2743 12.899 12 56.7
Portugal AA 19 -0.0742 17.556 AA 19 -0.0706 17.436 19 76.1
Romania B- 6 -0.3987 8.000 B- 6 -0.3942 8.324 6 31.2
Russia CCC- 3 -0.4428 5.161 SD 0 -0.4197 6.489 6 20
Singapore AAA 21 0.0073 19.716 AAA 21 0.0225 18.543 20 81.3
Slovak Republic BB+ 11 -0.2814 11.386 BB+ 11 -0.269 12.736 11 41.3
Slovenia A 16 -0.1922 14.319 A 16 -0.1878 15.275 15 58.4
South Africa BB+ 11 -0.2523 10.920 BB+ 11 -0.2386 12.873 12 45.8
Spain AA 19 -0.0924 17.433 AA+ 20 -0.0798 18.238 19 80.3
Sweden AA+ 20 0.0106 19.310 AA+ 20 0.0143 19.538 19 79.7
Switzerland AAA 21 0.071 23.436 AAA 21 0.0613 23.763 21 92.7
Thailand BBB- 12 -0.2452 11.207 BBB- 12 -0.2383 13.127 11 46.9
Trinidad & Tob BB+ 11 -0.2824 11.624 BBB- 12 -0.248 12.681 11 43.3
Tunisia BBB- 12 -0.2488 11.346 BBB- 12 -0.242 12.920 12 50.3
Turkey B 7 -0.4458 6.388 B 7 -0.4177 6.407 8 36.9
UK AAA 21 -0.0057 20.958 AAA 21 0.0062 20.769 21 90.2
United States AAA 21 0.0205 23.005 AAA 21 0.0264 23.981 21 92.2
Uruguay BBB- 12 -0.2695 12.516 BBB- 12 -0.2409 14.123 12 46.5
Venezuela B+ 8 -0.4444 6.999 B 7 -0.3921 8.422 7 34.4
RRR 08-2004 PAGE 1

Appendix: Correlation of difference matrices

Let us associate with any n-vector v its [n x n] “difference matrix” D whose component (i,j)
is the difference of the i’s and j’s components of v. In Section 4.4.2, we have used that the
correlation between any two n-vectors equals the correlation between their difference matrices.
In spite of the elementary nature of this statement, we could not find it in the literature, and shall
therefore provide here a formal proof of it.
Let us consider two n-dimensional vectors a and b, and two [n x n] dimensional difference
matrices C and D, the elements of which cij and dij are given by:
 cij = ai − a j ,
d = b − b , i, j = 1,..., n
 ij
(1.14)
i j

It can be seen from (1.14) that:

• C and D are anti-symmetric:
cij = − c ji and d ij = − c ji , i, j = 1,..., n, (1.15)

• C and D have diagonal elements equal to 0:

cii = d ii = 0, i = 1,..., n (1.16)
• the average of the elements of C and D is equal to 0:

c = ∑∑ cij / n 2 = 0 , d = ∑∑ d ij / n 2 = 0,
n n n n
(1.17)
i =1 j =1 i =1 j =1

We shall show that the correlation between the vectors a and b,

∑ (ai − a )(bi − b )
1 n
ρ ( a , b) =
n i =1

∑ i ∑ (bi − b )2
(1.18)
−
1 n 2 1 n
( a a )
n i =1 n i =1

is equal to the correlation between the matrices C and D

∑∑ (c − c )( d ij − d ) ∑∑ (a − a j )(bi − b j )
n n n n
1 1

ρ (C , D ) =
i =1 j =1 i =1 j =1
ij i
n2 n2

∑∑ (c ∑∑ (d ∑∑ (a ∑∑ (b − b )
= (1.19)
− c) − d) − aj)
n n n n n n n n
1 2 1 2 1 2 1 2

i =1 j =1 i =1 j =1 i =1 j =1 i =1 j =1
ij ij i i j
n2 n2 n2 n2

∑∑ (a − a j )2 can be rewritten as
n n
The expression
i =1 j =1
i

∑ ∑ [(a − a j )2 + ( a j − ai )2 ] = 2∑ ∑ ( ai − a j )2
n −1 n n −1 n
(1.20)
i =1 j = i +1 i =1 j = i +1
i

1
PAGE 2 RRR 08-2004

∑∑ (a − a j )(bi − b j ) can be rewritten as

n n
Also, the expression
i =1 j =1
i

∑ ∑ [(a − a j )(bi − b j ) + ( a j − ai )(b j − bi )] = 2∑ ∑ ( ai − a j )(bi − b j )

n −1 n n −1 n
(1.21)
i =1 j = i +1 i =1 j = i +1
i

Using (1.20) and (1.21), the correlation between the matrices C and D (1.10) can be rewritten
as :

∑ ∑ (a
n −1
− a j )(bi − b j )
n
1

ρ (C , D ) =
i =1 j = i +1
i
n2

∑ ∑ (a ∑ ∑ (b − b )
n −1 n −1
(1.22)
− aj)
n n
1 2 1 2

i =1 j = i +1 i =1 j = i +1
i i j
n2 n2

∑ (ai − a )2 in (1.18) can be rewritten as

1 n
Similarly, the expression
n i =1

∑ i n∑ − − = ∑∑ i j i − − = ∑∑(ai − a j )(ai − a )
1 n 1 n 1 n n 1 n n
( a a )( a a ) ( (a a ))( a a )
n i =1 j =1 n2 i =1 j =1 n2 i =1 j =1
j i

2 ∑∑
[(ai − a j )(ai − a ) +(a j − ai )(a j − a )] = 2 ∑ ∑ (ai − a j )(ai − a −a j + a ) = 2 ∑ ∑ (ai − a j )2 (1.23),
1 n−1 n 1 n−1 n 1 n−1 n
=
n i =1 j=i+1 n i=1 j=i+1 n i=1 j=i+1

∑ (ai − a )(bi − b ) in (1.18) can be rewritten as

1 n
while the expression
n i =1

∑ − ∑ a j )(bi − b ) = 2 ∑ ( ∑ (ai − a j ))(bi − b ) = ∑∑ (a − a j )( bi − b )

1 n 1 n 1 n n
1 n n
( a
n i =1 n j =1 i =1 j =1 i =1 j =1
i i
n n2

∑ ∑ [(a ∑ ∑ (a
n −1 n −1
= − a j )( bi − b ) + ( a j − ai )( b j − b )] = − a j )( bi − b − b j + b ) (1.24)
n n
1 1
i =1 j = i +1 i =1 j = i +1
i i
n2 n2

∑ ∑ (a
n −1
= − a j )(bi − b j )
n
1
i =1 j = i +1
i
n2

Using (1.23) and (1.24), the correlation between the vectors a and b (1.19) can be rewritten as

∑ ∑ (a
n −1
− a j )(bi − b j )
n
1

ρ ( a , b) =
i =1 j = i +1
i
n2

∑ ∑ (a ∑ ∑ (b − b )
n −1 n −1
, (1.25)
− aj )
n n
1 2 1 2

i =1 j =i +1 i =1 j = i +1
i i j
n2 n2

showing its equality to ρ (C , D ) . QED.